U.S. patent application number 16/383926 was filed with the patent office on 2020-07-30 for system and method for task-less mapping of brain activity.
This patent application is currently assigned to Washington University. The applicant listed for this patent is Maurizio Corbetta, Carl Hacker, Tim Laumann, Eric Leuthardt, Abraham Z. Snyder, Nicholas Szrama. Invention is credited to Maurizio Corbetta, Carl Hacker, Tim Laumann, Eric Leuthardt, Abraham Z. Snyder, Nicholas Szrama.
Application Number | 20200237316 16/383926 |
Document ID | / |
Family ID | 48281265 |
Filed Date | 2020-07-30 |
View All Diagrams
United States Patent
Application |
20200237316 |
Kind Code |
A9 |
Leuthardt; Eric ; et
al. |
July 30, 2020 |
SYSTEM AND METHOD FOR TASK-LESS MAPPING OF BRAIN ACTIVITY
Abstract
A computing device for use in a system for mapping brain
activity of a subject includes a processor. The processor is
programmed to select a plurality of measurements of brain activity
that is representative of at least one parameter of a brain of the
subject during a resting state. Moreover, the processor is
programmed to compare at least one data point from each of the
measurements with a corresponding data point from a previously
acquired data set from at least one other subject. The processor is
also programmed to produce at least one map for each of the
measurements based on the comparison of the resting state data
point and the corresponding previously acquired data point. The
processor may also be programmed to categorize the brain activity
in a plurality of networks in the brain based on the map.
Inventors: |
Leuthardt; Eric; (St. Louis,
MO) ; Szrama; Nicholas; (St. Louis, MO) ;
Hacker; Carl; (St. Louis, MO) ; Laumann; Tim;
(St. Louis, MO) ; Corbetta; Maurizio; (St. Louis,
MO) ; Snyder; Abraham Z.; (St. Louis, MO) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Leuthardt; Eric
Szrama; Nicholas
Hacker; Carl
Laumann; Tim
Corbetta; Maurizio
Snyder; Abraham Z. |
St. Louis
St. Louis
St. Louis
St. Louis
St. Louis
St. Louis |
MO
MO
MO
MO
MO
MO |
US
US
US
US
US
US |
|
|
Assignee: |
Washington University
St. Louis
MO
|
Prior
Publication: |
|
Document Identifier |
Publication Date |
|
US 20190239818 A1 |
August 8, 2019 |
|
|
Family ID: |
48281265 |
Appl. No.: |
16/383926 |
Filed: |
April 15, 2019 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
16136996 |
Sep 20, 2018 |
10258289 |
|
|
16383926 |
|
|
|
|
15237202 |
Aug 15, 2016 |
10092246 |
|
|
16136996 |
|
|
|
|
13673816 |
Nov 9, 2012 |
9480402 |
|
|
15237202 |
|
|
|
|
61558751 |
Nov 11, 2011 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
A61B 5/7246 20130101;
A61B 5/24 20210101; A61B 5/0022 20130101; A61B 2576/026 20130101;
G01R 33/5608 20130101; A61B 5/0042 20130101; A61B 5/055 20130101;
A61B 5/742 20130101; A61B 5/4064 20130101; G16H 50/20 20180101;
G01R 33/4806 20130101; A61B 5/0536 20130101 |
International
Class: |
A61B 5/00 20060101
A61B005/00; A61B 5/04 20060101 A61B005/04; A61B 5/053 20060101
A61B005/053; A61B 5/055 20060101 A61B005/055 |
Claims
1. A computer-implemented method for mapping a resting brain
activity of an individual subject at a plurality of positions
within the brain of the individual subject, the method comprising:
receiving, at a computing device, a plurality of correlation maps,
each correlation map comprising a plurality of elements, each
element comprising a correlation between one time-series
measurement and an additional time-series measurement selected from
of a plurality of time-series measurements, the one time-series
measurement obtained from one location within the brain of the
individual subject during a resting state, and the additional
time-series measurement obtained from one of a plurality of
additional locations within the brain of the individual subject
during a resting state; transforming, using the computing device,
each correlation map to a classification of brain activity using a
predetermined supervised classifier; and mapping, using the
computing device, each classification of brain activity transformed
from each correlation map to the one location within the brain of
the individual subject for each correlation map.
2. The method of claim 1, wherein transforming each correlation map
to a classification of brain activity further comprises using the
predetermined supervised classifier consisting of a perceptron, the
perceptron comprising a plurality of nodes arranged in an input
node layer, a hidden node layer, and an output node layer, the
perceptron characterized by the equation:
y.sub.o=.phi..sub.o[.SIGMA..sub.hw.sub.ho.phi..sub.h.SIGMA..sub.i(w.sub.i-
hy.sub.i))], wherein: y.sub.i is a vector of input node values of
the input node layer, each input node value corresponding to an
element of each correlation map; y.sub.o is a vector of output node
values of the output node layer, each output node value comprising
a probability of membership of each correlation map in each of a
plurality of resting state networks; w.sub.ih is a first
pre-determined weighting matrix configured to transform the vector
of input node values to a vector of hidden node input values;
.phi..sub.h is a pre-determined hidden node layer activation
function configured to transform the vector of hidden node input
values to a vector of hidden node values; w.sub.ho is a second
pre-determined weighting matrix configured to transform the vector
of hidden node values to a vector of output node input values; and
.phi..sub.o is a pre-determined output node layer activation
function configured to transform the vector of output node input
values to the vector of output node values.
3. The method of claim 2, further comprising: projecting, using the
computing device, each correlation map onto a plurality of
predetermined principal components to produce a vector in a
predetermined PCA space; and assigning, using the computing device,
the vector in the predetermined PCA space to the vector of input
node values.
4. The method of claim 3, further comprising, receiving, at the
computing device, a plurality of parameters defining the
predetermined supervised classifier, the plurality of parameters
comprising w.sub.ih, .phi..sub.h, w.sub.ho, and .phi..sub.o.
5. The method of claim 4, wherein .phi..sub.h is defined by the
equation: .phi..sub.h(v.sub.h)=atan h(bv.sub.h), wherein: v.sub.h
is the vector of hidden node input values, a is a first
predetermined constant, and b is a second predetermined
constant.
6. The method of claim 5, wherein .phi..sub.o is a logistic
activation function defined by the equation: .PHI. o ( v o ) = 1 1
+ e - av o , ##EQU00007## wherein: v.sub.o is the vector of output
node input values, and a is a predetermined constant.
7. The method of claim 6, wherein the input node layer comprises a
plurality of input nodes, each input node configured to receive one
value of the vector in the predetermined PCA space corresponding to
one of the plurality of principal components.
8. The method of claim 7, further comprising performing, using the
computing device, a principal components analysis on a training
data set to obtain the plurality of predetermined principal
components, the training data set comprising a plurality of
correlation maps corresponding to a plurality of positions within a
plurality of brains of a plurality of training subjects.
9. The method of claim 8, wherein the input node layer comprises a
plurality of input nodes configured to receive the vector of input
node values, wherein the number of input nodes ranges from about 10
to about 10,000.
10. The method of claim 9, wherein the hidden node layer comprises
a plurality of hidden nodes, each hidden node fully connected to
the plurality of input nodes by a plurality of weighted
feed-forward connections, the plurality of weighted feed-forward
connections characterized by the first pre-determined weighting
matrix, wherein the number of hidden nodes ranges from about 10 to
about 10,000.
11. The method of claim 10, wherein the output node layer comprises
a plurality of output nodes, each output node fully connected to
the plurality of hidden nodes by a second plurality of weighted
feed-forward connections, the second plurality of weighted
feed-forward connections characterized by the second pre-determined
weighting matrix, wherein the number of output nodes ranges from
about 5 to about 100.
12. The method of claim 11, wherein the plurality of output nodes
comprises a group of classification nodes and a nuisance node, each
classification node configured to receive the probability of
membership in one of the plurality of resting state networks and
the nuisance node configured to receive a probability of
classification as a non-active tissue selected from grey matter and
CSF.
13. The method of claim 12, wherein the plurality of resting state
networks comprise a dorsal attention network (DAN), a ventral
attention network (VAN), a sensorimotor network (SMN), a visual
network (VIS), fronto-parietal control network (FPC), a language
network (LAN), and a default mode network (DMN).
14. The method of claim 13, further comprising training, using the
computing device, the perceptron using a training data set, the
training data set comprising a plurality of training correlation
maps, each correlation training map further comprising a
pre-defined membership in one of the plurality of resting state
networks.
15. The method of claim 14, wherein training the perceptron further
comprises, for each training correlation map: transforming, using
the computing device, each training correlation map to an
intermediate classification of brain activity using the perceptron;
comparing, using the computing device, the intermediate
classification of brain activity with the pre-defined membership in
at least one of the plurality of resting state networks to
calculate a classification error; and updating, using the computing
device, at least one of the first pre-determined weighting matrix
and the second pre-determined weighting matrix to reduce the
classification error.
16. The method of claim 15, wherein comparing, using the computing
device, the intermediate classification of brain activity with the
pre-defined membership further comprises: calculating, using the
computing device, an error e.sub.o according to the equation:
e.sub.o(k)=d.sub.o(k)-y.sub.o(k), wherein: e.sub.o(k) is the
classification error obtained at a k.sup.th iteration of training,
d.sub.o(k) is the predefined membership in at least one of the
plurality of resting state networks, y.sub.o(k) is the intermediate
classification of brain activity obtained by the perceptron at the
k.sup.th iteration of training, and k is a training iteration
index; and calculating, using the computing device, a local
gradient .delta..sub.o of the error e.sub.o at the plurality of
output nodes according to the equation:
.delta..sub.o=e.sub.o.phi.'.sub.o(v.sub.o), wherein:
.phi.'.sub.o(v.sub.0) is a first derivative of the pre-determined
output node layer activation function.
17. The method of claim 16, wherein updating at least one of the
first pre-determined weighting matrix and the second pre-determined
weighting matrix further comprises: updating, using the computing
device, the second pre-determined weighting matrix according to the
equation:
w.sub.ho(k+1)=w.sub.ho(k)+.eta.(k).delta..sub.o(k)y.sub.h(k),
wherein: w.sub.ho(k+1) is the updated second pre-determined
weighting matrix, w.sub.ho(k) is an intermediate second
pre-determined weighting matrix, .eta.(k) is a learning rate, and
y.sub.h(k) is a vector of intermediate hidden node values of the
hidden node layer; and updating, using the computing device, the
second pre-determined weighting matrix according to the equation:
w.sub.ih(k+1)=w.sub.ih(k)+.eta.(k).delta..sub.o(k)y.sub.i(k):
wherein; w.sub.ih(k+1) is an updated first pre-determined weighting
matrix, w.sub.ih(k) is an intermediate first pre-determined
weighting matrix, and y.sub.i(k) is a vector of intermediate input
node values of the input node layer.
18. The method of claim 17, wherein the learning rate is an
adaptive learning rate configured to increase according to the
equation: .eta. ( k ) = A + K - A 1 + e - B ( log ( k ) - M ) ,
##EQU00008## wherein: A is an initial learning rate, B is an
exponential scaling factor, and K is a total number of training
iterations.
Description
CROSS REFERENCE TO RELATED APPLICATION
[0001] This application is a continuation of U.S. application Ser.
No. 15/237,202, filed Aug. 15, 2016, and incorporated herein in its
entirety. U.S. application Ser. No. 15/237,202 is a continuation of
U.S. application Ser. No. 13/673,816, filed Nov. 9, 2012, patented
as U.S. Pat. No. 9,480,402 and incorporated herein in its entirety.
U.S. application Ser. No. 13/673,816 claims priority to U.S.
Provisional Application Ser. No. 61/558,751, filed Nov. 11, 2011,
and also incorporated herein in its entirety.
BACKGROUND
[0002] The field of the invention relates generally to brain
mapping systems and, more particularly, to systems and methods for
task-less mapping of brain activity using resting state data
collected from a brain of a subject.
[0003] Brain mapping includes a set of neuroscience techniques that
are predicated on the mapping of biological quantities or
properties onto spatial representations of a subject's brain
resulting in at least one map. At least some known neuroimaging
systems or techniques are used frequently in clinical and research
settings for brain mapping such that brain function can be
monitored. For example, functional magnetic resonance imaging
(fMRI) may be used to enable researchers and clinicians to see
visual images of the brain, wherein the images may be used to
identify brain activity within a plurality of networks of the
brain. One approach includes a task based technique wherein the
fMRI may be used to detect correlations between brain activation
and various tasks that a subject performs during a scan. Such
task-based techniques can be useful in clinical applications. For
example, the images obtained through fMRI may enable a surgeon to
identify portions of the brain that are responsible for various
functions and the surgeon may attempt to avoid contact with such
portions while performing surgery on the brain.
[0004] However, task-based neuroimaging may not be suitable for all
segments of a clinical population. For example, a toddler or a
nervous patient may be unable to comprehend and/or perform various
tasks. It was recently discovered, via fMRI, that even during the
absence of overt tasks, fluctuations in brain activity are
correlated across functionally-related cortical regions. Thus, the
spatial and temporal evaluations of spontaneous neuronal activity
has allowed mapping of these resting-state networks (RSNs) with a
task-less technique. For this technique, at least one voxel within
the image obtained by the fMRI is selected and a correlation
analysis is performed to identify other voxels that correspond with
the selected voxel. However, it may be challenging to identify
which voxel to select. For example, an individual would require a
great deal of expertise and resources to select a relevant voxel.
In fact, selecting a voxel and/or processing information to select
a voxel can be time consuming.
[0005] Accordingly, it is desirable to provide a system and method
that can readily identify the voxels to select and, at
substantially the same time, provide suitable results for accurate
brain mapping.
BRIEF DESCRIPTION
[0006] In one aspect, a computing device for use in a system for
mapping brain activity of a subject generally comprises a
processor. The processor is programmed to select a plurality of
measurements of brain activity that is representative of at least
one parameter of a brain of the subject during a resting state.
Moreover, the processor is programmed to compare at least one data
point from each of the measurements with a corresponding data point
from a previously acquired data set from at least one other
subject. The processor is also programmed to produce at least one
map for each of the measurements based on the comparison of the
resting state data point and the corresponding previously acquired
data point. The processor may also be programmed to categorize the
brain activity in a plurality of networks in the brain based on the
map.
[0007] In another aspect, a system for mapping brain activity of a
subject generally comprises a sensing system and a computing device
that is coupled to the sensing system. The sensing system is
configured to detect a plurality of measurements of brain activity
that is representative of at least one parameter of a brain of the
subject during a resting state. The computing device includes a
communication interface that is configured to receive at least one
signal representative of the measurements, and a processor that is
coupled to the communication interface. The processor is programmed
to select the measurements of brain activity. Moreover, the
processor is programmed to compare at least one data point from
each of the measurements with a corresponding data point from a
previously acquired data set from at least one other subject. The
processor is also programmed to produce at least one map for each
of the measurements based on the comparison of the resting state
data point and the corresponding previously acquired data point.
The processor may also be programmed to categorize the brain
activity in a plurality of networks in the brain based on the
map.
[0008] In yet another aspect, a method for mapping brain activity
of a patient generally comprises selecting, via a processor, a
plurality of measurements of brain activity that is representative
of at least one parameter of a brain of the subject during a
resting state. At least one data point from each of the plurality
of measurements is compared, via the processor, with a
corresponding data point from a previously acquired data set from
at least one other subject. At least one map is produced, via the
processor, for each of the measurements based on the comparison of
the resting state data point and the corresponding previously
acquired data point. The brain activity is categorized, via the
processor, in a plurality of networks in the brain based on the
map.
BRIEF DESCRIPTION OF THE DRAWINGS
[0009] The patent or application file contains at least one drawing
executed in color. Copies of this patent or patent application
publication with color drawing(s) will be provided by the Office
upon request and payment of the necessary fee.
[0010] FIG. 1 is a block diagram of an exemplary system for
task-less mapping of brain activity;
[0011] FIG. 2 is a block diagram of an exemplary computing device
of the system shown in FIG. 1;
[0012] FIG. 3 is flow diagram of an exemplary method for task-less
mapping of brain activity using the system shown in FIG. 1;
[0013] FIG. 4 is an image of seed ROIs for generation of
correlation map data;
[0014] FIG. 5 is a schematic of an exemplary standard multi-layer
perceptron architecture and transfer function of the
perceptron;
[0015] FIG. 6 is a graph of a learning rate;
[0016] FIG. 7 is a schematic of a voxel-wise classification;
[0017] FIGS. 8A-8D are graphs depicting correlation maps;
[0018] FIGS. 9A-9F are graphs depicting performance levels;
[0019] FIG. 10 is a graph depicting search space for perceptron
architecture;
[0020] FIGS. 11A-11F are schematics of topographies in individual
subjects;
[0021] FIGS. 12A-12C are schematics of classification results;
[0022] FIG. 13 is a scan of group averaged results;
[0023] FIGS. 14A and 14B are graphs of exemplary evaluations;
[0024] FIG. 15 is a scan of voxels; and
[0025] FIGS. 16A and 16B are scans of topography estimates.
[0026] The patent or application file contains at least one drawing
executed in color. Copies of this patent or patent application
publication with color drawing(s) will be provided by the U.S.
Patent and Trademark Office upon request and payment of the
necessary fee.
DETAILED DESCRIPTION OF THE DRAWINGS
[0027] The exemplary systems, apparatus, and methods described
herein overcome at least some known disadvantages associated with
at least some known brain mapping techniques, such as task-based
and/or task-less systems. More specifically, the embodiments
described herein include a computing device for use in a system for
mapping brain activity of a subject that generally comprises a
processor. The processor is programmed to select a plurality of
measurements of brain activity that is representative of at least
one parameter of a brain of the subject during a resting state.
Moreover, the processor is programmed to compare at least one data
point from each of the measurements with a corresponding data point
from a previously acquired data set from at least one other
subject. The processor is also programmed to produce at least one
map for each of the measurements based on the comparison of the
resting state data point and the corresponding previously acquired
data point. The processor may also be programmed to categorize the
brain activity in a plurality of networks in the brain based on the
map. By using previously acquired data points to categorize the
brain activity in a plurality of networks in the brain of the
subject, task-based techniques can be avoided. Moreover, by having
the processor select the plurality of measurements, a user may no
longer need to spend a considerable amount of time determining
which measurements, such as voxels, to select.
[0028] FIG. 1 illustrates an exemplary system 100 for mapping brain
activity of a subject (not shown). It should be noted that the term
"brain activity" as used herein includes the various activities
within a brain of the subject that correspond to various tasks
performed by the subject. For example, the brain transmits and
receives signals in the form of hormones, nerve impulses, and
chemical messengers that enable the subject to move, eat, sleep,
and think. In the exemplary embodiment, system 100 is used to
identify locations within a plurality of networks within the brain
that are responsible for such brain activities.
[0029] As seen in FIG. 1, system 100 includes a sensing system 102
that is configured to detect a plurality of measurements of brain
activity that is representative of at least one parameter of the
brain of the subject during a resting state. In one suitable
embodiment, sensing system 102 is a magnetic resonance imaging
device (MRI) that is configured to generate at least one
spectroscopic signal representative of a plurality of measurements
of brain activity that is representative of at least one parameter
of the brain of the subject during a resting state. More
specifically, sensing system 102 may generate an altered magnetic
field within the brain to measure various parameters of the brain.
In another suitable embodiment, sensing system 102 may be a
specialized MRI, such as a functional magnetic resonance imaging
(fMRI) device that is used to measure a variation in blood flow
(hemodynamic response) related to neural activity in the brain or
spinal cord (not shown) of the subject. In yet another suitable
embodiment, sensing system 102 may be an electrocorticography
device having at least one electrode (not shown) to measure at
least one voltage fluctuation within the brain. It should be noted
that the present disclosure is not limited to any one particular
type of imaging and electrical technique or device, and one of
ordinary skill in the art will appreciate that the current
disclosure may be used in connection with any type of technique or
device that enables system 100 to function as described herein.
[0030] In the exemplary embodiment, system 100 also includes a
computing device 104 coupled to sensing system 102 via a data
conduit 106. It should be noted that, as used herein, the term
"couple" is not limited to a direct mechanical, electrical, and/or
communication connection between components, but may also include
an indirect mechanical, electrical, and/or communication connection
between multiple components. Sensing system 102 may communicate
with computing device 104 using a wired network connection (e.g.,
Ethernet or an optical fiber), a wireless communication means, such
as radio frequency (RF), e.g., FM radio and/or digital audio
broadcasting, an Institute of Electrical and Electronics Engineers
(IEEE.RTM.) 802.11 standard (e.g., 802.11(g) or 802.11(n)), the
Worldwide Interoperability for Microwave Access (WIMAX.RTM.)
standard, a short-range wireless communication channel such as
BLUETOOTH.RTM., a cellular phone technology (e.g., the Global
Standard for Mobile communication (GSM)), a satellite communication
link, and/or any other suitable communication means. IEEE is a
registered trademark of the Institute of Electrical and Electronics
Engineers, Inc., of New York, N.Y. WIMAX is a registered trademark
of WiMax Forum, of Beaverton, Oreg. BLUETOOTH is a registered
trademark of Bluetooth SIG, Inc. of Kirkland, Wash.
[0031] In the exemplary embodiment, computing device 104 is
configured to receive at least one signal representative of a
plurality of measurements of brain activity from sensing system
102. More specifically, computing device 104 is configured to
receive at least one signal representative of an altered magnetic
field within the brain of the subject from sensing system 102.
Alternatively, computing device 104 may be configured to receive at
least one signal representative of at least one voltage fluctuation
within the brain from at least one electrode.
[0032] System 100 also includes a data management system 108 that
is coupled to computing device 104 via a network 109. Data
management system 108 may be any device capable of accessing
network 109 including, without limitation, a desktop computer, a
laptop computer, or other web-based connectable equipment. More
specifically, in the exemplary embodiment, data management system
108 includes a database 110 that includes previously acquired data
of other subjects. In the exemplary embodiment, database 110 can be
fully or partially implemented in a cloud computing environment
such that data from the database is received from one or more
computers (not shown) within system 100 or remote from system 100.
In the exemplary embodiment, the previously acquired data of the
other subjects may include, for example, a plurality of
measurements of brain activity that is representative of at least
one parameter of a brain of each of the subjects during a resting
state. Database 110 can also include any additional information of
each of the subjects that enables system 100 to function as
described herein.
[0033] Data management system 108 may communicate with computing
device 104 using a wired network connection (e.g., Ethernet or an
optical fiber), a wireless communication means, such as, but not
limited to radio frequency (RF), e.g., FM radio and/or digital
audio broadcasting, an Institute of Electrical and Electronics
Engineers (IEEE.RTM.) 802.11 standard (e.g., 802.11(g) or
802.11(n)), the Worldwide Interoperability for Microwave Access
(WIMAX.RTM.) standard, a cellular phone technology (e.g., the
Global Standard for Mobile communication (GSM)), a satellite
communication link, and/or any other suitable communication means.
More specifically, in the exemplary embodiment, data management
system 108 transmits the data for the subjects to computing device
104. While the data is shown as being stored in database 110 within
data management system 108, it should be noted that the data of the
subjects may be stored in another system and/or device. For
example, computing device 104 may store the data therein.
[0034] During operation, while the subject is in a resting state,
sensing system 102 uses a magnetic field to align the magnetization
of some atoms in the brain of the subject and radio frequency
fields to systematically alter the alignment of this magnetization.
As such, rotating magnetic fields are produced and are detectable
by a scanner (not shown) within sensing system 102. More
specifically, in the exemplary embodiment, sensing system 102
detects a plurality of measurements of brain activity that is
representative of at least one parameter of the brain of the
subject during the resting state. Sensing system 102 also generates
at least one spectroscopic signal representative of the plurality
of measurements and transmits the signal(s) to computing device 104
via data conduit 106. Moreover, data of other subjects may be
transmitted to computing device 104 from database 110 via network
109. As explained in more detail below, computing device 104
produces at least one map, such as a functional connectivity map,
for each of the measurements based on a comparison of at least one
resting state data point of the subject and a corresponding data
point from the previously acquired data set from at least one other
subject. Computing device 104 uses the map to categorize or
classify the brain activity in a plurality of networks in the
brain.
[0035] FIG. 2 is a block diagram of computing device 104. In the
exemplary embodiment, computing device 104 includes a user
interface 204 that receives at least one input from a user, such as
an operator of sensing system 102 (shown in FIG. 1). User interface
204 may include a keyboard 206 that enables the user to input
pertinent information. User interface 204 may also include, for
example, a pointing device, a mouse, a stylus, a touch sensitive
panel (e.g., a touch pad, a touch screen), a gyroscope, an
accelerometer, a position detector, and/or an audio input interface
(e.g., including a microphone).
[0036] Moreover, in the exemplary embodiment, computing device 104
includes a presentation interface 207 that presents information,
such as input events and/or validation results, to the user.
Presentation interface 207 may also include a display adapter 208
that is coupled to at least one display device 210. More
specifically, in the exemplary embodiment, display device 210 may
be a visual display device, such as a cathode ray tube (CRT), a
liquid crystal display (LCD), an organic LED (OLED) display, and/or
an "electronic ink" display. Alternatively, presentation interface
207 may include an audio output device (e.g., an audio adapter
and/or a speaker) and/or a printer.
[0037] Computing device 104 also includes a processor 214 and a
memory device 218. Processor 214 is coupled to user interface 204,
presentation interface 207, and to memory device 218 via a system
bus 220. In the exemplary embodiment, processor 214 communicates
with the user, such as by prompting the user via presentation
interface 207 and/or by receiving user inputs via user interface
204. The term "processor" refers generally to any programmable
system including systems and microcontrollers, reduced instruction
set circuits (RISC), application specific integrated circuits
(ASIC), programmable logic circuits (PLC), and any other circuit or
processor capable of executing the functions described herein. The
above examples are exemplary only, and thus are not intended to
limit in any way the definition and/or meaning of the term
"processor."
[0038] In the exemplary embodiment, memory device 218 includes one
or more devices that enable information, such as executable
instructions and/or other data, to be stored and retrieved.
Moreover, memory device 218 includes one or more computer readable
media, such as, without limitation, dynamic random access memory
(DRAM), static random access memory (SRAM), a solid state disk,
and/or a hard disk. In the exemplary embodiment, memory device 218
stores, without limitation, application source code, application
object code, configuration data, additional input events,
application states, assertion statements, validation results,
and/or any other type of data. Computing device 104, in the
exemplary embodiment, may also include a communication interface
230 that is coupled to processor 214 via system bus 220. Moreover,
communication interface 230 is communicatively coupled to sensing
system 102 and to data management system 108 (shown in FIG. 1).
[0039] In the exemplary embodiment, processor 214 may be programmed
by encoding an operation using one or more executable instructions
and providing the executable instructions in memory device 218. In
the exemplary embodiment, processor 214 is programmed to select a
plurality of measurements that are received from sensing system 102
of brain activity that is representative of at least one parameter
of the brain of the subject during a resting state. The plurality
of measurements may include, for example, a plurality of voxels of
at least one image of the subject's brain, wherein the image may be
generated by processor 214 within computing device 104. The image
may also be generated by an imaging device (not shown) that may be
coupled to computing device 104 and sensing system 102, wherein the
imaging device may generate the image based on the data received
from sensing system 102 and then the imaging device may transmit
the image to computing device 104 for storage within memory device
218. Alternatively, the plurality of measurements may include any
other type measurement of brain activity that enables system 100 to
function as described herein.
[0040] Processor 214 may also be programmed to perform a
correlation analysis. More specifically, in the exemplary
embodiment, processor 214 may be programmed to compare at least one
data point from each of the plurality of measurements with a
corresponding data point from a previously acquired data set from
at least one other subject. For example, processor 214 may be
programmed to compare a resting state data point from each selected
voxel from an image of the subject with a corresponding data point
that is located within the same voxel of the previously acquired
data set of the other subject. Processor 214 may also be programmed
to produce at least one map (not shown in FIG. 2) of the brain of
the subject, such as a functional connectivity map, for each of the
plurality measurements. The map is based on the comparison of the
resting state data point and the corresponding previously acquired
data point. The map, for example, may illustrate the location
within the brain of a measured brain activity. Processor 214 may be
programmed to produce the map by using the various compared data
points in a known algorithm to calculate a plurality of outputs,
such as, for example, at least one output vector. One algorithm
that may be used is represented in Equation 1 below.
input 1 = tanh - 1 ( [ ln ( 1 / output - 1 ) - a ] pinv ( Weights
hidden - output ) a ) pinv ( Weights input - hidden ) b ( Eq . 1 )
##EQU00001##
[0041] In Equation 1, a and b represent activating function
parameters. The output represents a seven dimensional output vector
and pinv represents a pseudo inverse function.
[0042] Processor 214 may also be programmed to categorize or
classify the measured brain activity in a plurality of networks in
the brain based on the map. For example, processor 214 may be
programmed to categorize the measured brain activity to a
particular neural network of the brain of the subject based on the
location of the measured brain activity on the map of the subject's
brain.
[0043] During operation, as the subject is in a resting state,
sensing system 102 detects a plurality of measurements of brain
activity that is representative of at least one parameter of the
brain of the subject. Sensing system 102 transmits at least one
signal representative of the measurements to computing device 104
via data conduit 106. More specifically, the signals are
transmitted to and received by communication interface 230 within
computing device 104. Communication interface 230 then transmits
the signals to processor 214 for processing and/or to memory device
218, wherein the data may be stored and transmitted to processor
214 at a later time. Processor 214 may generate an image of the
plurality of measurements. Alternatively, sensing system 102 may
transmit the signals to an imaging device (not shown), wherein an
image of the measurements may be generated. The image may then be
transmitted to computing device 104, wherein the image is stored
within memory device 218 and transmitted to processor 214 for
processing.
[0044] Moreover, data of other subjects may be transmitted to
computing device 104 from database 110 (shown in FIG. 1) via
network 109 (shown in FIG. 1). More specifically, the data may be
received by communication interface 230 and then transmitted to
processor 214 for processing and/or to memory device 218, wherein
the data may be stored and transmitted to processor 214 at a later
time. Computing device 104 may obtain the data at any time during
operation.
[0045] In the exemplary embodiment, computing device 104 produces
at least one map for each of the plurality of measurements
received. More specifically, processor 214 first selects each of
the plurality of measurements, received from sensing system 102.
For example, in the exemplary embodiment, processor 214 selects
each of the voxels from the image. Alternatively, processor 214 may
select any other types of measurements for brain activity that
enables system 100 to function as described herein. Moreover, a
user may see the image on the computing device 104, via
presentation interface 207, and select the measurements, such as
voxels, via user interface 204.
[0046] When each of the measurements has been selected, processor
214 then performs a correlation analysis. More specifically,
processor 214 compares at least one data point from each of the
selected measurements with a corresponding data point from a
previously acquired data set from at least one other subject,
wherein computing device 104 obtained the data set from database
110. For example, processor 214 may compare at least one resting
state data point from each selected voxel of the image of the
subject with a data point that is located within the same voxel of
the previously acquired data set of at least one other subject.
[0047] When processor 214 has completed the correlation analysis,
processor 214 then produces at least one map (not shown in FIG. 2)
of the brain of the subject, such as a functional connectivity map,
for each of the measurements. More specifically, processor 214
produces a map of the brain of the subject based on each of the
comparisons of each of the resting state data points and the
corresponding previously acquired data points. The map, for
example, may illustrate the location within the brain of a measured
brain activity. Processor 214 then categorizes or classifies the
measured brain activity in a plurality of networks in the brain
based on the map. For example, based on the location of the
measured brain activity in the map, processor 214 categorizes the
measured brain activity to a particular neural network of the brain
of the subject. The map may be presented to the user via
presentation interface 207. Moreover, a textual representation
and/or a graphical output for the various categorizations may also
be presented to the user via presentation interface 207.
[0048] FIG. 3 is flow diagram of an exemplary method 300 for
task-less mapping of brain activity of a brain of a subject using
system 100 (shown in FIG. 1). A sensing system 102 (shown in FIG.
1) detects 302 a plurality of measurements of brain activity that
is representative of at least one parameter of the brain of the
subject during a resting state. Sensing system 102 transmits 304 at
least one signal representative of the measurements to a computing
device 104 (shown in FIGS. 1 and 2). The signals are received 306
by a communication interface 230 (shown in FIG. 2). The
measurements are selected 308 by a processor 214 (shown in FIG. 2).
At least one data point from each of the measurements is compared
310 with a corresponding data point from a previously acquired data
set from at least one other subject.
[0049] At least one map for each of the measurements is produced
312 based on the comparison of the resting state data point and the
corresponding previously acquired data point. The brain activity is
categorized 314 in a plurality of networks in the brain based on
the map. The map and/or an output for the categorization are
displayed 316 to a user, via a presentation interface 207 (shown in
FIG. 2).
[0050] The embodiments of the system and method for task-less
mapping of brain activity using resting state data of a brain of a
subject, as described herein, were used in the following exemplary
experiment.
Experiment
[0051] In the exemplary experiment, perceptron training and testing
used data sets previously acquired at the Neuroimaging Laboratories
at the Washington University School of Medicine in St. Louis, Mo.
All patients were young adults screened to exclude neurological
impairment and psychotropic medications. Demographic information
and acquisition parameters are given in Table 1 below.
TABLE-US-00001 TABLE 1 Characteristics of the training test and
validation data sets. Dataset Training Optimization (Test)
Validation Number of 21 (7M +14 F) 17 (8M + 9F) 10 (4M + 6F)
Subjects Age 27.6 (23-35) years 23.1 (18-27) years 23.3 .+-. 3
years Number of 128 .times. 6 runs 194 .times. 4 runs 100 .times. 9
runs frames TR (s) 2.16 2.16 3.03* *The TR in the validation data
set includes a one second pause between frames to accommodate
simultaneous EEG recording.
[0052] In the exemplary experiment, all imaging was performed with
a 3T Allegra scanner. Functional images were acquired using a BOLD
contrast sensitive gradient echo echo-planar sequence [FOV=256 mm,
flip angle=90.degree., 4 mm.sup.3 voxels, other parameters listed
in Table 1] during which subjects were instructed to fixate on a
visual cross-hair, remain still, and not fall asleep. Anatomical
imaging included one sagittal T1-weighted magnetization prepared
rapid gradient echo (MP-RAGE) scan (T1 W) and one T2-weighted scan
(T2 W).
[0053] Initial fMRI preprocessing followed conventional practice
known in the art. This included compensation for slice-dependent
time shifts, elimination of systematic odd-even slice intensity
differences due to interleaved acquisition, and rigid body
correction for head movement within and across runs. Atlas
transformation was achieved by composition of affine transforms
connecting the fMRI volumes with the T2 W and T1 W structural
images. Head movement correction was included with the atlas
transformation in a single resampling that generated volumetric
time series in 3 mm.sup.3 atlas space. Additional preprocessing in
preparation for correlation mapping included spatial smoothing (6
mm FWHM Gaussian blur in each direction), voxel-wise removal of
linear trends over each fMRI run, and temporal low-pass filtering
retaining frequencies below 0.1 Hz.
[0054] Spurious variance was reduced by regression of nuisance
waveforms derived from head motion correction and timeseries
extracted from regions (of "non-interest") in white matter and CSF.
Nuisance regressors included also the BOLD timeseries averaged over
the brain, i.e., global signal regression (GSR). Thus, all computed
correlations were effectively order 1 partial correlations
controlling for variance shared across the brain. GSR has been
criticized on the grounds that it artificially generates
anticorrelations. However, GSR fits well as a step preceding
principal component analysis because it generates approximately
zero-centered correlation distributions. As well, GSR enhances the
spatial specificity in subcortical seed regions and reduces
structured noise. The question of whether the left tail of a
zero-centered correlation distribution ("anticorrelations") is
"false" or "tenuously interpretable" is irrelevant in the context
of RSN classification.
[0055] Correlation maps were computed using standard seed-based
procedures, i.e., by correlating the timeseries averaged over all
voxels within the seed (generally, 5 mm spheres) against all other
voxels, excluding the first 5 (pre-magnetization steady-state)
frames of each fMRI run. Frame-censoring was employed with a
threshold of 0.5% RMS frame-to-frame intensity change.
Frame-censoring excluded 3.8.+-.1.1% of all magnetization
steady-state frames from the correlation mapping computations.
Correlation maps were Fisher z-transformed prior to further
analyses.
[0056] In the exemplary embodiment, Cortical reconstruction and
volume segmentation were performed using FreeSurfer. Adequate
segmentation was verified by inspection of the FreeSurfer-generated
results in each of the 21 training and 17 test datasets. Cortical
and subcortical gray matter regions were selected from these
segmentations, thresholded to obtain a conjunction of 30% of
subjects, and then masked with an image of the average BOLD signal
intensity across all subjects, thresholded at 80% of the mode
value. This last step removes from consideration brain areas in
which the BOLD signal is unreliable because of susceptibility
artifacts. The resulting 30,981 voxels constituted the grey matter
mask. This mask was applied to all correlation maps input to the
classifier. Individual surfaces were deformed to a common space,
producing consistent assignment of surface vertex indices with
respect to gyral features across subjects. Final volumetric results
for each subject were sampled onto surface vertices by cubic spline
interpolation onto mid-thickness cortical surface coordinates.
[0057] Seed regions were generated by meta-analyses of task-fMRI
studies. Task-response foci were initially assigned to one of 10
functional networks in Table 2 below. Each task fMRI study
contributed a variable number of foci (Task ROIs column in Table
2). Task foci were used as seeds to generate correlation maps in
all 21 subjects in the training set. These maps then were entered
into random effects analyses (against the null hypothesis of no
correlation) to produce Gaussianized t-statistic (Z-score) images.
Z-score images representing seeds assigned to the same RSN were
averaged. Additionally, a conjunction image representing at least
70% of random effects images for a given network (after
thresholding at |Z|>3) was produced. Averaged Z-score images
were masked to include only voxels contained in the conjunction.
Peaks of the conjunction-masked average were selected as center
coordinates for 6 mm spherical ROIs. Accordingly, the constraint
employed was that all ROIs within a given network must be separated
by at least 12 mm. This process resulted in a large set of ROIs
that were operationally treated as provisional.
TABLE-US-00002 TABLE 2 Studies of functional co-activation used to
generate seed ROIs. Pro- Final Task visional seed RSN Task paradigm
fMRI contrast ROIs seed ROIs ROIs DAN 1. Rapid Serial 1. Cue Type x
10 28 28 Visual Presentation event time (RSVP) 2. Cue location x 2.
Rapid Serial cue type x event Visual Presentation time (RSVP) 3.
Event time 3. Posner Cueing Task VAN Posner Cueing Task Invalid vs.
Valid 2 19 15 CO* Mixed design (10 Graph theoretic N/A 7 different
tasks) analysis* SMN Posner Cueing Task Target Period 11 37 39 AN
Various auditory Stimulation vs. 2 12 stimuli control VIS Visual
Localizer Peripheral 8 19 30 Foveal 2 12 FPC* Mixed design (10
Graph theoretic N/A 11 12 different tasks) analysis* LAN Perceptual
vs. Sentence 13 17 13 Episodic Memory Reading Search Paradigm DMN
Perceptual vs. Memory 4 42 32 Episodic Memory Retrieval Search
Paradigm *Regions reported were themselves the result of a
meta-analysis followed by refinement. Hence, these seeds were
directly used as provisional ROIs.
[0058] In the exemplary embodiment, the provisional ROI set was
iteratively refined by maximizing the spatial concordance between
the correlation map obtained from each seed and the map obtained by
pooling all seeds within the RSN to which the seed was assigned.
Pooled seed correlation maps were computed by averaging the time
series across all seeds assigned to each RSN. The single seed and
the pooled seed maps were averaged across subjects. RSN concordance
was assessed as the spatial correlation between the
(subject-averaged) single seed and the (subject-averaged) pooled
seed maps. Seeds were considered outliers if their concordance
estimate was less than 1.5 times the inter-quartile range below the
median of all other seeds in the RSN. Outlier seeds were reassigned
to the RSN of greatest concordance, unless they were maximally
concordant with the currently assigned RSN, in which case they were
removed entirely. After reassignment and outlier rejection, new
individual seed and pooled seed correlation maps were re-computed
and the process was iterated. Convergence (no reassignments or
outlier rejections) was achieved in 7 iterations. The
cingulo-opercular (CO) network did not survive iterative
refinement, and most seeds were reassigned to the ventral attention
network or removed. Similarly, the auditory network was subsumed
into the sensorimotor network and the originally distinct foveal
and peripheral visual networks were combined into a single (VIS)
network.
[0059] Iterative refinement yielded 169 ROIs representing 7 RSNs
with high intra- and low inter-network correlation, as shown in
FIGS. 4 and 15. To these were added a nuisance category consisting
of 6 ROIs in CSF spaces. The latter enabled the classifier to
separate correlation patterns representing CSF vs. true RSNs.
Computing correlation maps for each of the 175 seed regions in all
21 training subjects produced 3,675 images that were used as
training data. Each image in the training set was masked to include
only grey matter voxels, producing a 3,675.times.30,981 matrix.
Similarly, 17517=2,975 images were computed in the test data set.
Each image was assigned to one RSN (see the description of
iterative seed ROI refinement above and Table 2).
[0060] A multilayer perceptron was constructed to classify
resting-state fMRI correlation maps into 7 canonical spatial
patterns predefined as resting-state networks. The core of the
perceptron is an artificial neural network that includes an input,
hidden, and output layer, each consisting of some number of nodes
fully connecting to the next layer (all-to-all feed-forward).
Training samples (correlation maps from a particular seed and
subject) are passed into this feed-forward network and the output
is compared to the correct RSN label, as specified in the fMRI task
meta-analysis. The error in this comparison is used to update the
connections, or weights, between layers to increase the performance
of the classifier.
[0061] As an initial pre-processing step, the dimensionality of the
input data was reduced by using principal component analysis (PCA).
Representing correlation maps in terms of eigenvectors provides
efficient computation, well-conditioned weight matrices, and a free
parameter to represent the complexity of the input data (number of
PCs). PCA was performed on the matrix of masked correlation images
(21 subjects.times.175 seeds=3,675 images.times.30,981 voxels for
PCA). Each correlation map in the training (3,675 images) and the
test (2,975 images) data sets were then represented using a
variable number of principal components (PCs).
[0062] The input layer received the correlation map training data
as vectors in PCA space (the value of a given correlation map
projected along a particular PC). Thus, the number of input nodes
was a free parameter that depended on the number of PCs used to
represent the data. Each training example (a correlation map from a
particular seed ROI/subject pair) was associated with a desired
output value, d.sub.o (Eq. (7)), corresponding to the a priori RSN
labels. The goal of the training process is to compare the output
to these desired values, thereby generating an error signal used to
update connection weights. The overall transfer function of the
perceptron (Eq. (2)) corresponds to the detailed schematic of the
propagation of inputs through the perceptron (FIG. 5, see legend
for symbol definitions).
y o = .PHI. o ( h w ho ( .eta. ) .PHI. h ( i w ih y i ) ) ( 2 )
##EQU00002##
[0063] The total input to each hidden node, v.sub.h, is determined
by the sum of all input nodes, weighted by the feed-forward
connections (Eq. (3)). This sum is then transformed by the hidden
layer activation function to compute the output value of the hidden
layer node, y.sub.h (Eq. (4)).
v h = i w ih y i ( 3 ) y h = .PHI. h ( v h ) = a tanh ( b v h ) ( 4
) ##EQU00003##
[0064] The output layer nodes operate in the same manner as hidden
layer nodes (Eqs. (5) and (6)):
v o ( n ) = i w ho ( n ) y h ( n ) ( 5 ) y o = .PHI. o ( v o ) = 1
1 + e - a v o ( 6 ) ##EQU00004##
[0065] The After propagation of the input data through the
perceptron, the output value for each node, y.sub.o, was compared
to the desired value, d.sub.o, to find the error, e.sub.o (Eq.
(7)).
e.sub.o(k)=d.sub.o(k)-y.sub.o(k) (7)
[0066] The local gradient of the error at an output node is found
by the product of this error and the inverse of the activating
function applied to the output value:
.delta..sub.o=e.sub.o.phi.'.sub.o(v.sub.o) (8)
where the prime notation indicates the first derivative. After
every iteration (n), the weights for the hidden to output layer
connections were adjusted in the direction opposite of the gradient
of the error:
w.sub.ho(k+1)=w.sub.ho(k)+.eta.(k).delta..sub.o(k)y.sub.h(k)
(9)
where .eta. is the learning rate, y.sub.h is the value of hidden
layer node h, and .delta..sub.o is local error gradient at output
node o. Similarly, the weights to the hidden layer from the input
layer, w.sub.ih, are adjusted according to Eq. (10).
w.sub.ih(k+1)=w.sub.ih(k)+.eta..sub.h(k).delta..sub.h(k)y.sub.i(k)
(10)
[0067] The local gradient at a hidden node, .delta..sub.h, may be
computed by back-propagation from the output layer.
.delta. h = - dE d .gamma. h = .PHI. h ' ( v h ) o .delta. o w ho (
11 ) ##EQU00005##
[0068] The learning rate parameter, .eta., was set empirically. A
range of stable values was determined for a constant .eta., where
instability was noted as divergence or rapid oscillation of
classifier weights. The present results were obtained using an
adaptive learning rate that increased as a sigmoid in log iteration
index (FIG. 6). The initial learning rate was small
(.eta.(0)=A=510.sup.-4) to allow the classifier to begin a gentle
descent in error gradient towards a stable solution. The learning
rate increased exponentially (B=-3, Q=0.5), until saturating at an
empirically determined upper limit of stability to
(.eta.(.infin.)=210.sup.-3).
[0069] Separation of classes was quantified using receiver operator
characteristic analysis. Across a range of thresholds, the
proportion of within-class output values above the threshold (true
positive fraction, TPF) were compared to the number of out-of-class
values above the threshold (false positive fraction, FPF). The TPF
as a function of FPF defines the ROC curve. The area under the ROC
curve (AUC) was used as a summary statistic of classification
performance for each RSN class.
[0070] At logarithmically spaced intervals during the training
process, training was paused and AUCs were calculated in a separate
test data set. This procedure produced training trajectories
indicating the relative performance for each RSN (FIG. 9D)
throughout the training process. Peak performance for a given RSN
was defined at the iteration producing the maximum AUC value in the
test data (FIG. 9D). Overall performance was calculated as the
average of AUC values across networks (FIG. 9D).
[0071] In the exemplary embodiment, the number of PCs sampled
(N.sub.i), and the number of nodes in the hidden layer (N.sub.h)
constitute hyper-parameters subject to optimization. Overall RMS
error was evaluated over a densely sampled N.sub.i.di-elect
cons.[5, 6600].times.N.sub.h.di-elect cons.[4, 5000] space. For
each (N.sub.i, N.sub.h) coordinate, a classifier was trained until
test set error reached a minimum. The architecture with the least
error (minimum of eight repetitions for each coordinate) was
selected (FIG. 10).
[0072] After identifying the architecture with least error in the
test data set, performance was further optimized by simulated
annealing, countering the tendency of perceptrons to become trapped
in local minima. Mimicking the random movement of atoms aligning in
cooling metal, simulated annealing uses random perturbations of
model parameters to find the global extremum in an objective
function. Perturbations of steadily decreasing size (specified by a
`cooling profile`) are guaranteed to find a global minimum with
slow enough cooling, although, in practice, the necessary cooling
profile is prohibitively slow. After training the perceptron until
a minimum in RMS test set error, every weight, {w.sub.ih} and
{w.sub.ho}, was multiplied by a random coefficient. Training was
then resumed to find a new minimum. If lower error was achieved,
the new weights were accepted. This process was then repeated.
[0073] The value for each weight was determined by first sampling
from a uniform distribution, x.di-elect cons.[-1, 1], transformed
by a hyperbolic function, N=(1-x)/(1+x). Thus each weight was
multiplied by N.di-elect cons.[0,.infin.], and was thus unchanged
when x=0. The range sampled within x determined the amount of noise
injected into the system, using values closer to zero over the
course of cooling. The maximum value of x was determined by the
temperature, T, and the minimum value was determined so that the
mean squared value of N was unity:
1 T - a .intg. a T ( x - 1 x + 1 ) 2 dx = 1 ( 12 ) ##EQU00006##
[0074] This choice of noise ensured that the sum of squares of the
connection weights was unaltered by perturbation and that most
weights were decreased, while a small selection was sporadically
increased. A geometric cooling function (Eq. (13)) was used, which
decreased over K.sub.1 perturbation epochs; this entire annealing
process was repeated K.sub.2 times, each time with a slightly
cooler temperature profile.
T.sub.k.sub.1.sub.,k.sub.2=T.sub.0r.sup.(k.sup.1.sup.+3k.sup.2.sup.)
(13)
[0075] The following parameters were used: r=0.95, T.sub.0=0.4,
K.sub.1=40, K.sub.2=20. To map RSNs in individual subjects, a
correlation map was generated for every voxel in the brain and then
classified using the trained and optimized perceptron. An overall
schematic of this process is depicted in FIG. 10. A correlation map
was produced for every point in the brain by correlating every
voxel's BOLD time-course with every other voxel in the brain. Each
map was masked before classification to include only grey matter
voxels producing a 65,549 (voxels within brain mask).times.30,981
(voxels within grey matter mask) element matrix (FIG. 7). This data
was then projected onto the eigenvectors of the training data,
reducing the dimensionality to 65,549.times.2500 (FIG. 7). Thus,
all correlation maps were represented in the same input data space
for classifier training and testing. The reduced whole-brain
connectivity data was then propagated through the perceptron, with
the first layer reducing the data to 22 features (65,549.times.22;
FIG. 7), and the second layer producing RSN estimates
(65,549.times.8, FIG. 7). However, FIG. 7 depicts only 7 output
classes because one of the 8 outputs is a nuisance component used
only in post-processing.
[0076] Classifier output values are approximately uniformly [0,1]
distributed as a result of the logistic activation function on the
output layer (Eq. (6)). Classifier values were then normalized
within each voxel to sum to unity (FIG. 7). This normalization
penalized voxels that had high classification values for multiple
networks. The presence of a CSF classification component further
penalized RSN estimates in voxels exhibiting CSF-related
correlation patterns. Within each network, classifier values were
then converted back to an exactly uniform [0,1] distribution across
voxels (rank-order transform). This transformation resulted in
voxels ranked in membership for each network across the brain
expressed as a percentile.
[0077] To visualize group level results on the cortical surface,
RSN topography estimates were projected to the cortical
mid-thickness surface for each subject (after surface-registration
across subjects). Averages were then computed across surface nodes.
The standard deviation of classifier values was also calculated
node-wise to illustrate regions of high variability. These
group-level results were projected onto the group-average inflated
surface. To visualize group level results in sub-cortical
structures, classifier values were averaged voxel-wise across
subjects. Group-average images were then re-sampled to 1 mm cubic
voxels and overlaid on a co-registered MNI152 atlas target.
[0078] In the exemplary embodiment, spatial correlation analysis
(FIG. 8B) and principal component analysis (FIG. 8C) of the
training data (the correlation maps produced for each seed ROI)
revealed distinct clustering corresponding to RSNs. In the
map-to-map spatial correlation matrix (averaged across subjects),
training inputs showed high correlation with other inputs of the
same RSN compared to inputs of other classes (FIG. 8B).
Additionally, the map-to-map correlation matrix showed two major
clusters, one corresponding to the DAN, VAN, VIS, and MOT networks,
and the other corresponding to the FPC, LAN, and DMN networks.
Projection of all 3,675 correlation maps into principal component
space gave rise to partially overlapping clusters corresponding to
7 RSNs. In the PC1.times.PC2 plane (FIG. 6), DAN (purple) and DMN
(red) showed little overlap and appeared at opposite ends of the
PC1 axis. MOT (light blue) and VIS (green) clusters were highly
overlapping in this plane, but showed little overlap in the
PC3.times.PC4 plane.
[0079] FIGS. 9A-9F shows the training performance for the
perceptron optimized for overall performance (2500 input PCs, 22
hidden layer nodes). For every correlation map, the perceptron
output node values represent an estimate of membership for each
RSN. The expectation value of all initial perceptron outputs is 0.5
(FIGS. 9A-9C) as the expected output value with zero-mean weights
(v.sub.0) is 0 (Eqs. (5) and (6)). As training progresses,
within-class output values increase towards unity (e.g., DMN output
node values for DMN inputs, red traces in FIG. 9A), while
out-of-class output values decrease towards zero (DMN output value
for non-DMN ROI-derived maps, all other traces in FIG. 9A).
[0080] Area under the ROC curve (AUC) trajectories are shown in
FIG. 9D. This quantity, averaged across RSN classes, began near
chance (0.65 after one iteration) and rose in later iterations. For
all networks, the AUC exhibited a transient decrement in
performance early in training. This feature corresponded to
transient changes of slope in RMS error but did not produce
concavity (local minima) in RMS error (FIG. 9E). Class separation
was achieved at varying numbers of iterations for different RSNs.
Across all perceptron architectures, the default mode network (red
trace) always achieved asymptotic performance earliest, and the
language network (orange) latest. Asymptotic performance for CSF
classification occurred much later than any true RSN. Performance
on the test data initially followed training performance until
reaching a global maximum (FIG. 9E). This maximum occurred at
varying iteration indices for different RSNs. Training beyond this
point resulted in over-fitting, manifesting as decreasing test data
performance despite increasing training performance. Inputs that
were previously correctly classified in the test data became
incorrectly classified (FIGS. 9B and 9C).
[0081] Over a dense sampling of input and hidden layer sizes (N ix
N.sub.h), the perceptron was trained until the peak AUC could be
determined (FIG. 10). The optimal overall performance for the
perceptron was found at 2500 PCs and 22 hidden layer nodes (FIG.
10). The perceptron was trained with this architecture using 10 mm
ROIs and the result was optimized through simulated annealing,
yielding an over-all classification performance of 0.9822 (AUC)
with 17.1% RMS error. The maximal AUC and minimal RMS error rates
differed by network, as shown in Table 3.
TABLE-US-00003 TABLE 3 RSN classification performance. Test Retest
(Optimization Set) (Validation Set) Accuracy Error Accuracy Error
Network (AUC) (RMS) (AUC) (RMS) DAN 0.973 20.2% 0.973 20.1% VAN
0.971 17.9% 0.979 17.6% SMN 0.988 16.4% 0.994 17.2% VIS 0.993 13.4%
0.998 12.7% FPC 0.972 17.5% 0.989 14.8% LAN 0.985 14.9% 0.991 14.4%
DMN 0.993 14.4% 0.990 17.6% Mean 0.982 17.1% 0.988 16.6%
[0082] These values reflect MLP training with 10 mm radius seeds
(FIGS. 14A and 14B) and optimization with simulated annealing.
After completion of classifier training, voxel-wise connectivity
patterns were classified in individual subjects in the test data
set (FIG. 11A). RSN topography summaries were computed as
winner-take-all maps (FIGS. 11A-11F and 12A-12C). Well-defined RSN
topography was obtained in all subjects in the test and training
groups. Specifically, the subject-wise mean and standard error of
the AUC was 0.982.+-.0.007, with the worst performing subject at
0.963. These figures corresponded to RMS error of 16.5.+-.1.4% with
the worst subject yielding 19.1%. RSNs were generally contiguous
regions that conformed to previously described topography. The
relationship of perceptron-defined RSNs to previous findings is
discussed more fully below.
[0083] RSN topography estimates were averaged over all subjects in
the training and test groups. FIGS. 12A-12C addresses both the
central tendency (top row) of each group as well as inter-subject
variability (middle row). Average network topographies had higher
values near locations of ROIs used to generate training maps. This
is expected because voxels within ROIs are likely to have similar
correlation maps to the ROI. High classification values (in the top
25%) were also found in contiguous regions not used to generate
training data. For example, a lateral temporal region was
classified as a fronto-parietal control region, and a dorsal
pre-motor region was classified into the language network. This
type of result demonstrates external validity (or equivalently,
generalizability) of perceptron classification, i.e., recovery of
true features in RSNs not included in the training set. These
features are also present in the results in individual subjects
(FIG. 11A).
[0084] Further evidence of external validity is shown in FIG. 13.
For example, thalamic voxels approximately corresponding to nucleus
ventralis posterior were classified as SMN, substantially in
agreement. Similarly, voxels in the posterior cerebellum (Crus I
and II) and the cerebellar tonsils were classified as DMN (FIG. 13,
Z=-30, Z=-47), substantially in agreement. These results are
notable because neither cerebellar nor thalamic ROIs were used to
generate training data. Further, no cerebellar voxels were within
the grey matter mask, which means that the classifier successfully
identified cerebellar ROIs purely on the basis of cortical
connectivity maps. The perceptron generated asymmetric
classification in the cerebellum for the VAN and LAN networks
(Z=-30), contralateral to their asymmetric cerebral representation
(Z=+47).
[0085] The present results (individual and group RSN topographies)
exhibit a high degree of face validity with respect to the training
data and previously reported RSN results (FIGS. 11A-11F, 12A-12C,
and 13). Thus, for example, components of the DMN used as seeds to
generate the training data were classified as DMN in all subjects.
This was true not only for easily classified networks (e.g., the
DMN) but also for networks (e.g., VAN and LAN) that are
inconsistently found by unsupervised procedures. The results shown
in FIGS. 11A-11F illustrate that the perceptron reliably classified
RSNs in each test set individual (AUC>0.971), even in cases in
which the RMS error was relatively high (>0.2).
[0086] However, inter-individual differences were also evident
(FIGS. 11A-11F). These differences systematically varied according
to RSN and exhibited RSN-specific zones of high as well as low
inter-individual variability (FIGS. 12A-12C). Easily classified
voxels (i.e., with high classification values) generally showed the
least inter-subject variability. Such regions, e.g., the posterior
parietal component of the DMN (FIGS. 12A-12C), were surrounded by
zones of high variability (e.g., ring around the angular gyms). The
pre- and post-central gyri consistently showed high SMN
classification values but were bordered by regions of high
inter-subject SMN variability. Interestingly, inter-subject
variability was low also in areas with classification values near
0, particularly in areas typically anticorrelated with other
networks (e.g., low DAN variance in the angular gyrus, a component
of the DMN; low DMN variance in MT+, a component of the DAN).
[0087] At least four factors potentially contribute to observed
inter-subject classifier output variability: (i) limited or
compromised fMRI data, (ii) limitations intrinsic to the MLP (iii)
true inter-individual differences in RSN topography and (iv)
misregistration. Each of the possibilities is discussed below.
[0088] With regard to (i), the fMRI data used in the present work
were obtained in healthy, cooperative young adults. Hence, the
fraction of frames excluded because of head motion was low (about
4%). The total quantity of fMRI data acquired in each individual
was, by current standards, generous. However, fMRI data quantity
clearly affects MLP performance (see 4.4.2 below and FIGS.
14A-14B). Current results suggest that more data generally improves
MLP performance. The impact of fMRI data quality and quantity on
MLP performance in clinical applications remains to be determined.
With regard to (ii), the observation of zones with high classifier
values and low variance bordered by regions of high variance may
reflect classification uncertainty in areas that truly represent
more than one RSN, i.e., voxels with high participation
coefficients.
[0089] With regard to (iii), on the other hand, the presently
observed inter-individual differences may truly reflect individual
variability in RSN topography. Previous work has demonstrated that
inter-individual differences in task-evoked activity correspond to
"transition zones" in resting state networks (e.g., the boundary
between parietal DMN and DAN regions). These same regions appear in
our inter-subject variance maps for both DMN and DAN (FIGS.
12A-12C). We also note that areas of high RSN classification
variability (pre-frontal, parietal, lateral temporal) broadly
correspond to regions exhibiting the greatest expansion over the
course of human development and evolution. This correspondence may
possibly be coincidental, but it is consistent with the hypothesis
that later developing or evolutionarily more recent areas of the
brain tend to be more variable across individuals.
[0090] With regard to (iv), some proportion of the variability in
observed RSN topography estimates may be explained by uncorrected
anatomical variability. To investigate this possibility, the
overall RSN standard deviation map (FIG. 16A) was compared to
sulcal depth variability (FIG. 16B) and a weak spatial correlation
(r=0.2) was found. By inspection, these maps were concordant only
at a broad spatial scale: both showed low variability in primary
motor/auditory/insular cortices and high variability elsewhere.
Little correspondence was evident at finer scales (note lack of
annular patterns in FIG. 16B). The degree to which anatomical
variability contributes to spurious variance in RSN topography
estimates may be addressed by measuring the degree to which
non-linear or surface-based registration decreases inter-subject
variance and increases overall classifier performance (higher AUC,
lower RMSE).
[0091] Two distinct types of external validity, that is, correct
classification outside the training set, are evident in our
results. First, high overall classification performance was
achieved for a priori seed-based correlation maps in test (98.2%
AUC) and hold-out datasets (98.8% AUC). Performance was reliable in
all subjects (97.1% worst-case AUC), which is critical in clinical
applications. Second, and perhaps of greater scientific interest,
the RSN estimates in areas not covered by seed regions were
strongly concordant with previously reported task-based and
resting-state fMRI results. For example, while no temporal FPC seed
ROI was included in the training set, a posterior temporal gyrus
locus was classified as FPC the group level (FIGS. 12A-12C).
Similarly, the MLP also identified the parahippocampal gyrus as DMN
and a dorsal pre-motor region that has been associated with
articulation of speech as LAN. The right inferior cerebellum was
first associated with language function by PET studies of semantic
association tasks. Identification of this region here as part of
the LAN network (FIG. 13, WTA, Z<-30) is doubly significant.
First, no cerebellar seeds were used to generate training data and,
further, cerebellar voxels were excluded from the gray matter mask,
hence, were not seen by the classifier. Second, lateralized RSN
components typically are not found by unsupervised seed-based
correlation mapping.
[0092] These findings highlight the capabilities of supervised
classifiers applied to the problem of identifying RSNs in
individuals. The representation of language (primarily Broca's and
Wemicke's areas) has been extensively studied using task-based fMRI
and correlation mapping with a priori selected ROIs. However, the
language network, as presently defined, typically is not recovered
as such by unsupervised methods. Rather, components of the LAN are
generally found only at fine-scale RSN descriptions. Thus, an RSN
including Broca's and Wemicke's areas appears as the 11th of 23
components in; these same areas were identified as VAN and DMN. A
component consistent with the presently defined LAN at a
hierarchical level of 11 (but not 7) clusters has been found. Thus,
the exemplary experiment, work demonstrates the potential of
supervised classifiers to find networks that are subtle features of
the BOLD correlation structure, possibly even minor sub-components
within hierarchically organized RSNs, that nevertheless have high
scientific and/or clinical value. The LAN was specifically included
here to meet the clinical imperative of localizing language
function in the context of pre-operative neurosurgical
planning.
[0093] In the exemplary embodiment, the hierarchical scale of an
RSN is reflected in training performance trajectories (FIGS. 9E and
F): in all (N.sub.i.times.N.sub.h) architecture variants, the DMN
was the first to be separated from other RSNs. The DMN arguably is
the most robust feature in the correlation structure of intrinsic
brain activity. Its topography is very similar across RSN mapping
strategies (specifically, spatial ICA and seed-based correlation
mapping. Here, the DMN and regions anticorrelated with the DMN were
well separated along the first principal component of the training
data (FIG. 6).
[0094] After the DMN, the sensorimotor and visual networks were
next to achieve separation during classifier training. These
networks are often seen at the next level down in the RSN hierarchy
as offshoots of the anti-DMN or extrinsic system. The dorsal
attention network achieved only a small peak in error descent
compared to other `extrinsic` networks, though this occurred in
close proximity (note overlap of DAN, MOT, VIS peaks in FIG. 9F).
In contrast, the LAN and VAN were last to achieve separation during
training. This corresponds to the observation that LAN and VAN
systems are typically found by analyses extending to lower levels
of the RSN hierarchy.
[0095] In the exemplary embodiment, the observer is a multi-layer
perceptron and the task is to assign RSN labels to each voxel.
Performance is evaluated in terms of mean squared classification
error and ROC analyses. It follows that MLP performance can be used
to evaluate image quality across a wide range of variables, e.g.,
scanners, and acquisition parameters (e.g., TR, run length,
resolution), preprocessing strategies (nuisance regression,
filtering, spatial smoothing) and data representations (surface or
volume based). This principle is demonstrated by systematically
evaluating MLP performance in relation to quantity of fMRI data and
seed ROI size.
[0096] The relation between total quantity of fMRI data and MLP
performance (test dataset RMS error) is shown in FIG. 14A. The
plotted points represent five replicate MLP training/test runs. RMS
error as a function of data quantity was well fit (R.sup.2=0.994)
by a three-parameter empirically derived hyperbolic function. The
parameterized function implies that classifier error monotonically
decreases with increasing total fMRI data length but ultimately
asymptotes at .about.18% RMS (with 5 mm radius seeds and no
simulated annealing). The existence of this asymptote may indicate
that resting-state brain networks are inherently non-separable in
the sense of classification. This is consistent with the notion of
"near decomposability" of hierarchical systems formed by multiple,
sparsely inter-connected modules. This concept has since been
extended to brain networks.
[0097] The relationship between seed ROI radius and RMS
classification error was explored using a perceptron architecture
optimized with 5 mm radius seeds (2500 PCs, 22 hidden nodes). All
seeds were masked to include only gray matter voxels. The results
of systematically varying seed ROI size are shown in FIG. 14B. A
clear minimum in RMS error was obtained with seeds of approximately
10 mm radius. Voxel-wise RSN topographies were qualitatively
similar across ROI sizes, but larger seeds generated less noisy
RSNs with more pronounced peaks. This result is unexpected, as it
deviates from the current standard practice of using approximately
6 mm radius seeds. There are several possible explanations for the
present results. Large seeds may best match the characteristic
dimensions of RSNs in the 7-network level description of the brain.
Alternatively, large seeds may compensate for misregistration in
affine-coregistered, volume-preprocessed data. Smaller seeds may be
used in classifiers operating on non-linear, surface-coregistered,
geodesically smoothed data. The results shown in FIG. 14B reflect
the effect of seed radius on the correlation maps used to train the
MLP. It is formally possible for a corrupted training set to yield
a better classifier as evaluated by test set classification error.
Thus, the results shown in FIG. 14B should not be interpreted as
unambiguously indicating that 10 mm radius seeds are optimal for
correlation mapping.
[0098] Inter-individual differences in computed RSN topographies
may reflect multiple factors. Cross-gyral contamination due to the
relatively large voxels used in this study (4 mm acquisition, 3 mm
post-processing analyses) may limit the precision of RSN
classification in the dataset. Potential strategies for validating
perceptron-derived results include comparison with measures of
structural (axonal) connectivity and invasive electrophysiologic
recording.
[0099] The MLP RSN classifier operates at the voxel level via
computed correlation maps. After training, it reliably identifies
RSN topographies in individual subjects. Classification is rapid (2
minutes using Matlab running on Intel i7 processors) and automated,
hence suitable for deployment in clinical environments. After
training, classification is independent of any particular seed.
Therefore, the trained MLP is expected to be robust to anatomical
shifts and distortions, for example, owing to enlarged ventricles
and mass effects or even loss of neural tissue (e.g., stroke).
[0100] In this experiment, the classifier was trained to operate in
3D image space for compatibility with clinical imaging formats.
However, the MLP concept can be readily adapted to operate on
correlation maps represented on the cortical surface. Similarly,
voxel-wise classifiers can be trained to classify subjects despite
anatomical abnormalities (e.g., brain tumors) by altering the
domain of the training set, i.e., excluding tumor voxels. Another
potentially useful MLP modification would be removal by regression
of the relationship between correlation and distance to the seed.
Such regression may decrease the reliance of the classifier on
local connectivity, thereby reducing susceptibility to corruption
by movement artifact.
[0101] As compared to known systems that are used for brain
mapping, the embodiments described herein enable a substantially
efficient task-less system for brain mapping. More specifically,
the embodiments described herein include a computing device for use
in a system for mapping brain activity of a subject that generally
comprises a processor. The processor is programmed to select a
plurality of measurements of brain activity that is representative
of at least one parameter of a brain of the subject during a
resting state. Moreover, the processor is programmed to compare at
least one data point from each of the measurements with a
corresponding data point from a previously acquired data set from
at least one other subject. The processor is also programmed to
produce at least one map for each of the measurements based on the
comparison of the resting state data point and the corresponding
previously acquired data point. The processor may also be
programmed to categorize the brain activity in a plurality of
networks in the brain based on the map. By using previously
acquired data points to categorize the brain activity in a
plurality if networks in the brain of the subject, task-based
techniques may be avoided. Moreover, by having the processor select
the plurality of measurements, a user may no longer need to spend a
considerable amount of time determining which measurements, such as
voxels, to select.
[0102] Exemplary embodiments of the system, apparatus, and method
are described above in detail. The system, apparatus, and method
are not limited to the specific embodiments described herein, but
rather, components of the system and apparatus, and/or steps of the
methods may be utilized independently and separately from other
components and/or steps described herein. For example, but not
limited to, the system may also be used in combination with other
apparatus, systems, and methods, and is not limited to practice
with only the system as described herein. Rather, the exemplary
embodiment can be implemented and utilized in connection with many
other applications.
[0103] Although specific features of various embodiments of the
invention may be shown in some drawings and not in others, this is
for convenience only. In accordance with the principles of the
invention, any feature of a drawing may be referenced and/or
claimed in combination with any feature of any other drawing.
[0104] Although described in connection with an exemplary computing
system environment, embodiments of the invention are operational
with numerous other general purpose or special purpose computing
system environments or configurations. The computing system
environment is not intended to suggest any limitation as to the
scope of use or functionality of any aspect of the invention.
[0105] Embodiments of the invention may be described in the general
context of computer-executable instructions, such as program
modules, executed by one or more computers or other devices. The
computer-executable instructions may be organized into one or more
computer-executable components or modules. Generally, program
modules include, but are not limited to, routines, programs,
objects, components, and data structures that perform particular
tasks or implement particular abstract data types. Aspects of the
invention may be implemented with any number and organization of
such components or modules. For example, aspects of the invention
are not limited to the specific computer-executable instructions or
the specific components or modules illustrated in the figures and
described herein. Other embodiments of the invention may include
different computer-executable instructions or components having
more or less functionality than illustrated and described herein.
Aspects of the invention may also be practiced in distributed
computing environments where tasks are performed by remote
processing devices that are linked through a communications
network. In a distributed computing environment, program modules
may be located in both local and remote computer storage media
including memory storage devices.
[0106] In operation, a computer executes computer-executable
instructions embodied in one or more computer-executable components
stored on one or more computer-readable media to implement aspects
of the invention described and/or illustrated herein.
[0107] The order of execution or performance of the operations in
embodiments of the invention illustrated and described herein is
not essential, unless otherwise specified. That is, the operations
may be performed in any order, unless otherwise specified, and
embodiments of the invention may include additional or fewer
operations than those disclosed herein. For example, it is
contemplated that executing or performing a particular operation
before, contemporaneously with, or after another operation is
within the scope of aspects of the invention.
[0108] When introducing elements of aspects of the invention or the
embodiments thereof, the articles "a," "an," "the," and "said" are
intended to mean that there are one or more of the elements. The
terms "comprising," "including," and "having" are intended to be
inclusive and mean that there may be additional elements other than
the listed elements.
[0109] This written description uses examples to disclose the
invention, including the best mode, and also to enable any person
skilled in the art to practice the invention, including making and
using any devices or systems and performing any incorporated
methods. The patentable scope of the invention is defined by the
claims, and may include other examples that occur to those skilled
in the art. Such other examples are intended to be within the scope
of the claims if they have structural elements that do not differ
from the literal language of the claims, or if they include
equivalent structural elements with insubstantial differences from
the literal language of the claims.
* * * * *