U.S. patent application number 13/781134 was filed with the patent office on 2013-09-19 for magnetic resonance spectroscopy of breast biopsy to determine pathology, vascularization and nodal development.
The applicant listed for this patent is Carolyn E. Mountford, Peter Russell, Ian C.P. Smith, Rajmund L. Somorjai. Invention is credited to Carolyn E. Mountford, Peter Russell, Ian C.P. Smith, Rajmund L. Somorjai.
Application Number | 20130245957 13/781134 |
Document ID | / |
Family ID | 33518710 |
Filed Date | 2013-09-19 |
United States Patent
Application |
20130245957 |
Kind Code |
A1 |
Mountford; Carolyn E. ; et
al. |
September 19, 2013 |
Magnetic Resonance Spectroscopy of Breast Biopsy to Determine
Pathology, Vascularization and Nodal Development
Abstract
Robust classification methods analyse magnetic resonance
spectroscopy (MRS) data (spectra) of fine needle aspirates taken
from breast tumours. The resultant data when compared with the
histopathology and clinical criteria provide computerized
classification-based diagnosis and prognosis with a very high
degree of accuracy and reliability. Diagnostic correlation
performed between the spectra and standard synoptic pathology
findings contain detail regarding the pathology (malignant versus
benign), vascular invasion by the primary cancer and lymph node
involvement of the excised axillary lymph nodes. The classification
strategy consisted of three stages: pre-processing of MR magnitude
spectra to identify optimal spectral regions, cross-validated
Linear Discriminant Analysis, and classification aggregation via
Computerised Consensus Diagnosis. Malignant tissue was
distinguished from benign lesions with an overall accuracy of 93%.
From the same spectrum, lymph node involvement was predicted with
an accuracy of 95% and tumour vascularisation with an overall
accuracy of 92%.
Inventors: |
Mountford; Carolyn E.; (East
Ryde, AU) ; Russell; Peter; (Pyme, AU) ;
Smith; Ian C.P.; (Winnipeg, CA) ; Somorjai; Rajmund
L.; (Headingly, CA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Mountford; Carolyn E.
Russell; Peter
Smith; Ian C.P.
Somorjai; Rajmund L. |
East Ryde
Pyme
Winnipeg
Headingly |
|
AU
AU
CA
CA |
|
|
Family ID: |
33518710 |
Appl. No.: |
13/781134 |
Filed: |
February 28, 2013 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
12072327 |
Feb 26, 2008 |
8404487 |
|
|
13781134 |
|
|
|
|
11012959 |
Dec 15, 2004 |
7335511 |
|
|
12072327 |
|
|
|
|
09961776 |
Sep 24, 2001 |
|
|
|
11012959 |
|
|
|
|
Current U.S.
Class: |
702/19 ;
324/300 |
Current CPC
Class: |
G16H 10/40 20180101;
G01R 33/4625 20130101; G01R 33/465 20130101; Y10T 436/24 20150115;
G01R 33/20 20130101; G01N 24/08 20130101; A61P 43/00 20180101 |
Class at
Publication: |
702/19 ;
324/300 |
International
Class: |
G06F 19/00 20060101
G06F019/00; G01R 33/20 20060101 G01R033/20 |
Goverment Interests
STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT
[0002] The work described herein was supported by U.S. Army Grant
number DAMD 17-96-1-6077 and NH & MRC 950215 and NH & MRC
973769.
Claims
1.-2. (canceled)
3. A method of determining whether a primary breast tumor in breast
tissue is accompanied by a tumor which has invaded a vascular
region outside the breast tissue, comprising the steps of:
obtaining magnetic resonance spectra of only a primary breast tumor
in breast tissue; comparing the spectra obtained from only primary
breast tumors in breast tissue with reference spectra which
contains (i) first reference values from only primary breast tumors
for primary breast tumors for which a tumor has not invaded a
vascular region outside the breast tissue and (ii) second different
reference values from only primary breast tumors for which a tumor
is also present in a vascular region outside the breast tissue; and
determining whether a tumor has invaded a vascular region outside
the breast tissue based on the comparison of the obtained spectra
from only the primary breast tumor with the first and second
reference values.
4. A reference data structure containing spectral data only from
primary breast tumors for which a tumor is not present in any
vascular region outside the breast tissue and different spectral
data only from primary breast tumors for which a tumor is present
in at least one vascular region outside the breast tissue.
5. A method of determining whether a primary tumor in primary
tissue is accompanied by a tumor in at least one vascular region
outside the primary tissue, comprising the steps of: obtaining
magnetic resonance spectra of only a primary tumor in primary
tissue, comparing the spectra obtained from only the primary tumor
with reference spectra which contains (i) first reference values
from only primary tumors for which a tumor is not present in any
vascular region outside the primary tissue and (ii) second
different reference values from only the primary tumors for which a
tumor is present in at least one vascular region outside the
primary tissue; and determining whether a tumor is present in at
least one vascular region outside the primary tissue based on the
comparison of the obtained spectra from only the primary tumor with
the first and second reference values.
6. A reference data structure containing spectral data only from
primary tumors in primary tissue for which a tumor is not present
in any vascular region outside the primary tissue and different
spectral data only from primary tumors in primary tissue for which
a tumor is present in at least one vascular region outside the
primary tissue.
7. A computer readable storage medium which stores a program for
determining whether a primary tumor in primary tissue is
accompanied by a tumor in at least one vascular region outside the
primary tissue, the program comprising the steps of: obtaining
magnetic resonance spectra of a primary tumor in primary tissue;
comparing the spectra obtained from only the primary tumor with
reference spectra which contains (i) first reference values from
only primary tumors for primary tumors in primary tissue for which
a tumor is not present in any vascular region outside the primary
tissue and (ii) second different reference values from only the
primary tumors for which a tumor is also present in at least one
vascular region outside the primary tissue; and determining whether
a tumor is present in at least one vascular region outside the
primary tissue based on the comparison of the obtained spectra only
from the primary tumor with the first and second reference
values.
8. A computer readable storage medium which stores a reference data
structure containing first spectral data only from primary tumors
in primary tissue for which a tumor is not present in any vascular
region outside the primary tumor and different spectral data only
from primary tumors in primary tissue for which a tumor is present
in at least one vascular region outside the primary tumor.
9. A method for obtaining a statistical classifier for classifying
spectral data from a biopsy of tissue to determine the
classification of a characteristic of the tissue, comprising: (a)
locating a plurality of discriminatory subregions in magnetic
resonance spectra of biopsies of tissue having known classifiers of
a characteristic, (b) cross-validating the spectra by selecting a
first portion of the spectra comprising a first plurality of the
spectra leaving the remainder of the spectra, developing linear
discriminant analysis classifiers from said first portion of
spectra, and validating the remainder of the spectra using the
classifiers from the first portion of the spectra, to obtain
optimized linear discriminant analysis coefficients, (c) repeating
step (b) a plurality of times, each time selecting a different
portion of the spectra to form the first portion, to obtain a
different set of optimized linear discriminant analysis
coefficients for each of said plurality of times; and (d) obtaining
an average of the linear discriminant analysis coefficients to
obtain final classifier spectra indicating the classification of
the characteristic based on the spectra, wherein spectra from a
biopsy of tissue of unknown classification of a characteristic may
be compared to the final classifier spectra to determine the
classification of the characteristic of the tissue.
10. An apparatus for obtaining a statistical classifier for
classifying spectral data from a biopsy of tissue to determine the
classification of a characteristic of the tissue, comprising: (a) a
locator for locating a plurality of discriminatory subregions in
magnetic resonance spectra of biopsies of tissue having known
classifiers of a characteristic of tissue, (b) a crossvalidator for
selecting a first portion of the spectra comprising a first
plurality of the spectra leaving the remainder of the spectra,
developing linear discriminant analysis classifiers from said first
portion of spectra, and validating the remainder of the spectra
using the classifiers from the first portion of the spectra to form
the first portion, to obtain optimized linear discriminant analysis
coefficients, said cross-validator selecting, developing and
validating a plurality of times, each time selecting a different
portion of the spectra, to obtain a different set of optimized
linear discriminant analysis coefficients for each of said
plurality of times, and (c) an averager for obtaining an average of
the linear discriminant analysis coefficients to obtain final
classifier spectra indicating the classification of the
characteristic based on the spectra, whereby spectra from a biopsy
of tissue of unknown classification of a characteristic may be
compared to the final classifier spectra to determine the
classification of the characteristic of the tissue.
11. A method for determining the classification of a characteristic
of tissue, comprising: obtaining magnetic resonance spectra of a
biopsy of tissue having unknown classification of a characteristic
and comparing the spectra with a classifier, said classifier having
been obtained by: (a) locating a plurality of discriminatory
subregions in the magnetic resonance spectra of biopsies of tissue
having known classifications of a characteristic of the tissue, (b)
cross-validating the spectra of (a) by selecting a first portion of
spectra comprising a first plurality of the spectra leaving the
remainder of the spectra, developing linear discriminant analysis
classifier from said first portion of spectra, and validating the
remainder of the spectra using the classifications from the first
portion of the spectra, to obtain optimized linear discriminant
analysis coefficients, (c) repeating step (b) a plurality of times,
each time selecting a different portion of the spectra to form the
first portion, to obtain a different set of optimized linear
discriminant analysis coefficients for each of said plurality of
times, and (d) obtaining a weighted average of the linear
discriminant analysis coefficients to obtain final classifier
spectra indicating the classification of the characteristic based
on the spectra, wherein the spectra from the biopsy of tissue
having unknown classification may be compared to the final
classifier spectra to determine the classification of the
characteristic of the tissue.
12. An apparatus for determining the classification of a
characteristic of tissue, comprising: a spectrometer for obtaining
magnetic resonance spectra of a biopsy of tissue having unknown
classification of a characteristic; a classifier for statistically
classifying the spectra by comparing the spectra with a reference
classifications, said classifier having been obtained by: (a)
locating a plurality of discriminatory subregions in the magnetic
resonance spectra of biopsies of tissue having known
classifications of a characteristic of the tissue, (b)
cross-validating the spectra of (a) by selecting a first portion of
spectra comprising a first plurality of the spectra leaving the
remainder of the spectra, developing linear discriminant analysis
classifier from said first portion of spectra, and validating the
remainder of the spectra using the classifiers from the first
portion of the spectra, to obtain optimized linear discriminant
analysis coefficients, (c) repeating step (b) a plurality of times,
each time selecting a different portion of the spectra to form the
first portion, to obtain a different set of optimized linear
discriminant analysis coefficients for each of said plurality of
times, and (d) obtaining an average of the linear discriminant
analysis coefficients to obtain final classifier spectra indicating
the classification of the characteristic based on the spectra, and
wherein said classifier compares the spectra from the biopsy of
tissue having unknown classification to the final classifier
spectra to determine the classification of the characteristic of
the tissue.
Description
CROSS REFERENCES TO RELATED APPLICATIONS
[0001] This application claims priority on, and incorporates by
reference, U.S. Provisional Application Ser. No. 60/160,029 filed
Oct. 18, 1999.
BACKGROUND OF THE INVENTION
[0003] 1. Technical Field of the Invention
[0004] The present invention relates to the use of magnetic
resonance spectroscopy, and more particularly to such use for
determining pathology, vascularization and nodel involvement of a
biopsy of breast tissue.
[0005] 2. Description of the Related Art
[0006] Within this application several publications are referenced
by arabic numerals within parentheses. Full citations for these and
other references may be found at the end of the specification
immediately preceding the claims. The disclosures of all of these
publications in their entireties are hereby incorporated by
reference into this application in order to more fully describe the
state of the art to which this invention pertains.
[0007] Clinical evaluation, mammography and aspiration cytology or
core biopsy (triple assessment) is undertaken on women presenting
with breast lesions in most Western countries. Clinical assessment
of palpable breast lumps is unreliable (1, 2). Impalpable lesions
are usually discovered by screening or diagnostic mammography,
which has a reported sensitivity of 77-94% and a specificity of
92-95% (3). Cytological assessment of fine needle aspiration
biopsies (FNAB) has sensitivities ranging from 65-98% and
specificities ranging from 34-100% (4) depending on the skill of
the person performing the aspiration and the expertise of the
cytopathologist.
[0008] Following surgical excision of the lesion a time consuming
process of preparation and pathological assessment of the specimen
determines the nature of the tumour and the prognostic features
associated with it.
SUMMARY OF THE INVENTION
[0009] Magnetic resonance spectroscopy (MRS) is a modality with a
proven record in the diagnosis of minimally invasive malignant
lesions (5-11). MR spectra of small samples of tissue or even cell
suspensions enable the reliable determination of whether the tissue
of origin is malignant or benign. Often MRS is able to detect
malignancy before morphological manifestations are visible by light
microscopy (8).
[0010] The potential of proton MRS from FNAB specimens to
distinguish benign from malignant breast lesions has been
demonstrated previously (12). At that time the MRS method relied on
visual reading to process spectra and calculate the ratio of the
diagnostic metabolites choline and creatine. This spectral ratio
allowed tissue to be identified as either benign or malignant. In a
small cohort of 20 patients within that study it also distinguished
high grade ductal carcinoma in situ (DCIS) with comedonecrosis or
microinvasion from low grade DCIS. Despite the limitation of visual
inspection, which could only assess those spectra with a signal to
noise ratio (SNR) of greater than 10, the visual method resulted in
a diagnosis of malignant or benign with a sensitivity and
specificity of 95 and 96%. FIG. 1 shows malignant and benign
spectra with good SNR while FIG. 2 shows spectra with poor SNR.
[0011] Twenty percent of the spectra were discarded because low
aspirate cellularity yielded inadequate SNR. In the initial study
visual analysis used only two of fifty or more available resonances
(6). Thus, potentially diagnostic and prognostic information in the
remaining spectrum may have been ignored.
[0012] A 3-stage, robust statistical classification strategy (SCS)
has been developed to classify biomedical data and to assess the
full MR spectrum obtained from biological samples. The robustness
of the method has been demonstrated previously with the analysis of
proton MR spectra of thyroid tumours (13), ovarian (14), prostate
(9), and brain tissues (15). The present invention applies SCS to
assess proton MR spectra of breast aspirates against pathological
criteria in order to determine the correct pathology on samples
with sub-optimal cellularity and SNR and to determine if other
diagnostic and prognostic information is available in the
spectra.
[0013] The inventors have determined that SCS on MRS from breast
FNAB is more reliable than visual inspection to determine whether a
lesion is benign or malignant, and that a greater proportion of
spectra is useful for analysis. Furthermore, spectral information
obtained from MRS on FNAB of breast cancer specimens predicted
lymph node metastases (overall accuracy of 96%) and vascular
invasion (overall accuracy of 92%).
[0014] The invention provides a method for obtaining a statistical
classifier for classifying spectral data from a biopsy of breast
tissue to determine the classification of a characteristic of the
breast tissue, comprising: [0015] (a) locating a plurality of
maximally discriminatory subregions in magnetic resonance spectra
of biopsies of breast tissue having known classifiers of a
characteristic, [0016] (b) cross-validating the spectra by
selecting a portion of the spectra, developing linear discriminant
analysis classifiers from said first portion of spectra, and
validating the remainder of the spectra using the classifiers from
the first portion of the spectra, to obtain optimized linear
discriminant analysis coefficients, [0017] (c) repeating step (b) a
plurality of times, each time selecting a different portion of the
spectra, to obtain a different set of optimized linear discriminant
analysis coefficients for each of said plurality of times; [0018]
(d) obtaining a weighted average of the linear discriminant
analysis coefficients to obtain final classifier spectra indicating
the classification of the characteristic based on the spectra; and
[0019] (e) comparing spectra from a biopsy of breast tissue of
unknown classification of a characteristic to the final classifier
spectra to determine the classification of the characteristic of
the breast tissue.
[0020] The invention provides an apparatus for obtaining a
statistical classifier for classifying spectral data from a biopsy
of breast tissue to determine the classification of a
characteristic of the breast tissue, comprising: [0021] (a) a
locator for locating a plurality of maximally discriminatory
subregions in magnetic resonance spectra of biopsies of breast
tissue having known classifiers of a characteristic of breast
tissue, [0022] (b) a cross-validator for selecting a portion of the
spectra, developing linear discriminant analysis classifiers from
said first portion of spectra, and validating the remainder of the
spectra using the classifiers from the first portion of the
spectra, to obtain optimized linear discriminant analysis
coefficients, said cross-validator selecting, developing and
validating a plurality of times, each time selecting a different
portion of the spectra, to obtain a different set of optimized
linear discriminant analysis coefficients for each of said
plurality of times, and [0023] (c) an averager for obtaining a
weighted average of the linear discriminant analysis coefficients
to obtain final classifier spectra indicating the classification of
the characteristic based on the spectra, whereby spectra from a
biopsy of breast tissue of unknown classification of a
characteristic may be compared to the final classifier spectra to
determine the classification of the characteristic of the breast
tissue.
[0024] The invention provides a method for determining the
classification of a characteristic of breast tissue, comprising:
[0025] obtaining magnetic resonance spectra of a biopsy of breast
tissue having unknown classification of a characteristic and
comparing the spectra with a classifier, said classifier having
been obtained by: [0026] (a) locating a plurality of maximally
discriminatory subregions in the magnetic resonance spectra of
biopsies of breast tissue having known classifications of a
characteristic of the breast tissue, [0027] (b) cross-validating
the spectra of (a) by selecting a portion of spectra, developing
linear discriminant analysis classifier from said first portion of
spectra, and validating the remainder of the spectra using the
classifications from the first portion of the spectra, to obtain
optimized linear discriminant analysis coefficients, [0028] (c)
repeating step (b) a plurality of times, each time selecting a
different portion of the spectra, to obtain a different set of
optimized linear discriminant analysis coefficients for each of
said plurality of times, and [0029] (d) obtaining a weighted
average of the linear discriminant analysis
[0030] coefficients to obtain final classifier spectra indicating
the classification of the characteristic based on the spectra,
and
comparing the spectra from the biopsy of breast tissue having
unknown classification to the final
[0031] The invention provides an apparatus for determining the
classification of a characteristic of breast tissue, comprising:
[0032] a spectrometer for obtaining magnetic resonance spectra of a
biopsy of breast tissue having unknown classification of a
characteristic; [0033] a classifier for statistically classifying
the spectra by comparing the spectra with a reference
classifications, said classifier having been obtained by: [0034]
(a) locating a plurality of maximally discriminatory subregions in
the magnetic resonance spectra of biopsies of breast tissue having
known classifications of a characteristic of the breast tissue,
[0035] (b) cross-validating the spectra of (a) by selecting a
portion of spectra, developing linear discriminant analysis
classifier from said first portion of spectra, and validating the
remainder of the spectra using the classifiers from the first
portion of the spectra, to obtain optimized linear discriminant
analysis coefficients, [0036] (c) repeating step (b) a plurality of
times, each time selecting a different portion of the spectra, to
obtain a different set of optimized linear discriminant analysis
coefficients for each of said plurality of times, and [0037] (d)
obtaining a weighted average of the linear discriminant analysis
coefficients to obtain final classifier spectra indicating the
classification of the characteristic based on the spectra, and
wherein said classifier compares the spectra from the biopsy of
breast tissue having unknown classification to the final classifier
spectra to determine the classification of the characteristic of
the breast tissue.
DESCRIPTION OF THE DRAWINGS
[0038] FIG. 1 shows malignant and benign spectra with relatively
good SNR;
[0039] FIG. 2 shows spectra with relatively poor SNR; and
[0040] FIG. 3 shows a system for determining pathology,
vascularization and nodal involvement according to the
invention.
DETAILED DESCRIPTION OF THE INVENTION
Methods
Preparation of Patients:
[0041] Intra-operative FNAB were taken from 139 patients undergoing
breast surgery for malignant and benign conditions (Table 1) by
three surgeons in separate hospitals. In order to provide a
sufficiently large data set for SCS an additional 27 patients
joined the study (see Table 1). Impalpable breast lesions that had
been localised by carbon track or hook wire were included except if
the lesion was not palpable at excision or when the pathology
specimen could have been compromised. All samples were taken during
surgery under direct vision after the lesion had been identified
and incised sufficiently widely to ensure that the FNAB and tissue
specimens represented the same lesion and were thus comparable. The
lesion was identified and incised in-vivo via the margin with the
greatest apparent depth of normal tissue between it and the lesion
to ensure the pathologist could report upon the lesion according to
a standard protocol. Malignant and suspicious lesions were
orientated with sutures and radio opaque vascular clips (Ligaclips)
for pathological and radiological orientation. The FNAB was
collected by the surgeon using a 23-gauge needle on a 5 ml syringe.
The number of needle passes was recorded and the surgeon's
evaluation of the quality of the aspirate was made. Before the
needle was removed from the lesion, a tissue sample including the
relevant part of the needle track was taken. The size of this
tissue specimen was estimated and recorded by the surgeon.
[0042] Pre-operative clinical and investigative data included
localised pain, nipple discharge or nipple crusting, details of
previous mammography, and whether the lesion was detected through
screening. The clinical, mammography, ultra sonographic,
cytological, core biopsy and MRI details were recorded as
malignant, suspicious, benign, impalpable, uncertain or not
done.
[0043] The pathology specimen was sent on ice at the initial
stages, but later in formalin, for standard histopathological
reporting and hormone receptor analysis. The pathology report was
issued in synoptic format (16).
[0044] Specific tumour-related clinical and sampling information
was collected. These included a history of previous breast biopsies
with dates, diagnoses, sizes and sites of these lesions along with
the current lesion's duration, palpability, laterality, size and
locality within the breast. The date of operation, the extent of
breast surgery from open biopsy to total mastectomy, and axillary
surgery from sampling to level 3 dissection was recorded.
Specimen Preparation:
[0045] Following complete excision of the lesions the FNAB cytology
and tissue specimens were placed in polypropylene vials containing
300 ml phosphate-buffered saline (PBS) in D.sub.2O. All specimens
were immediately immersed in liquid nitrogen and stored at
-70.degree. C. for up to 6 weeks until MRS analysis.
[0046] Prior to the proton MRS experiment, each FNAB specimen was
thawed and transferred directly to a 5 mm MRS tube. The volume was
adjusted to 300 ml with PBS/D.sub.2O where necessary. Proton MRS
assessment of all specimens was performed without knowledge of the
correlative histopathology, either from the synoptic pathology
report or from sectioning of tissue used in MRS study.
[0047] The sample of tissue excised around the needle tract was
similarly placed in polypropylene vials containing 300 ml
PBS/D.sub.2O and immersed in liquid nitrogen as described above.
This sample was later used for pathological correlation.
Data Acquisition:
[0048] MRS experiments were carried out on a Broker Avance 360
wide-bore spectrometer (operating at 8.5 Tesla) equipped with a
standard 5 mm dedicated proton probehead. The sample was spun at 20
Hz and the temperature maintained at 37.degree. C. The residual
water signal was suppressed by selective gated irradiation. The
chemical shifts of resonances were referenced to aqueous sodium
3-(trimethylsilyl)-propanesulphonate (TSPS) at 0.00 ppm.
One-dimensional spectra were acquired over a spectral width of 3597
Hz (10.0 ppm) using a 90.degree. pulse of 6.5-7 .mu.s, 8192 data
points, 256 accumulations and a relaxation delay of 200 seconds,
resulting in a pulse repetition time of 3.14 seconds.
[0049] SNR was determined using the Bruker standard software
(xwinnmr). The noise region was defined between 8.5-9.5 ppm. The
signal region was defined between 2.8 to 3.5 ppm.
Histopathology:
[0050] Diagnostic correlation was obtained by comparing spectral
analysis with the hospital pathology report provided for each
patient. Lymph node involvement and vascular invasion were
determined from the reports only in cases where this information
was complete. In the participating hospitals lymph nodes were
embedded and serial sectioned in standard fashion. One 5 .mu.m
section out of every 50 (i.e., each 250 .mu.m) was stained and
examined. All intervening sections were discarded.
[0051] In the initial phase of the study, cytological analysis of
the aspirate after MRS analysis was attempted but cellular detail
was compromised by autolytic changes and this approach was not
pursued. In order to verify FNAB sampling accuracy, a separate
histopathological assessment by a single pathologist (PR) was
obtained from tissue removed from the aspiration site of the MRS
sample. Tissue specimens were thawed, fixed in FAA (formalin/acetic
acid/alcohol), paraffin-embedded, sectioned at 7 .mu.m, stained
with haematoxylin and eosin according to standard protocols and
reviewed under the light microscope by the pathologist without
access to the clinical or MRS data. Tissue preservation, abundance
of epithelial cells relative to stroma, and resence of potentially
confounding factors such as fat and inflammatory cells were
reported in addition to the principal diagnosis.
Statistical Classification Strategy:
[0052] The general classification strategy has been developed and
was designed specifically for MR and IR spectra of the biofluids
and biopsies. The strategy consists of three stages. First the MR
magnitude spectra are preprocessed, (in order to eliminate
redundant information and/or noise) by submitting them to a
powerful genetic algorithm-based Optimal Region Selection (ORS_GA)
(17), which finds a few (at most 5-10) maximally discriminatory
subregions in the spectra. The spectral averages in these
subregions are the ultimate features and used at the second stage.
This stage uses the features found by ORS_GA to develop Linear
Discriminant Analysis (LDA) classifiers that are made robust by
IBD's bootstrap-based crossvalidation method (18). The
crossvalidation approach proceeds by randomly selecting about half
the spectra from each class and using these to train a classifier
(usually LDA). The resulting classifier is then used to validate
the remaining half. This process is repeated B times (with random
replacement), and the optimized LDA coefficients are saved. The
weighted average of these B sets of coefficients produces the final
classifier. The ultimate classifier is the weighted output of the
500-1000 different bootstrap classifier coefficient sets and was
designed to be used in a clinical setting as the single best
classifier. The classifier consists of probabilities of class
assignment for the individual spectra. For 2-class problems, class
assignment is called crisp if the class probability is >0.75%.
For particularly difficult classification problems the third stage
is activated. This aggregates the outputs (class probabilities) of
several independent classifiers to form a Computerised Consensus
Diagnosis (CCD) (13, 15). The consequence of CCD is that
classification accuracy and reliability is generally better than
the best of the individual classifiers.
[0053] FIG. 3 shows a spectrometer 10, which may be a Bruker Avance
360 spectrometer operating at 8.5 Tesla, with equipped computer.
The statistical classification strategy (SCS) computer 12 stores
the SCS and other programs described herein. The clinical data base
includes the information from the data acquisition and
histopathology, used by the computer 12 to develop the classifier
16. The classifier 16 classifies the characteristics (e.g.
pathology, vascularization and/or lymph node involvement) of the
breast tissue under examination.
Results
[0054] One hundred and sixty-six patients were involved in the
study. A summary of the clinicopathological criteria is shown in
Table 1.
Benign Versus Malignant:
[0055] Proton MR spectra were recorded for each FNAB irrespective
of the cellularity of the aspirate. However, those specimens with a
SNR less than 10, which were shown to be inadequate for visual
inspection (12) have been included in the SCS analysis without
significantly compromising accuracy. Visual inspection of all
spectra irrespective of signal to noise gave a sensitivity and
specificity of 85.3% and 81.5% respectively (Table 2a), based on
the creatine-to-choline ratio. When SCS-based classifiers were
developed for all available spectra (Table 2b), 96% of the spectra
were considered crisp and could be assigned unambiguously by the
classifier as malignant or benign. Sensitivity and specificity were
93% and 92% respectively.
[0056] After removing the 31 spectra with the previously determined
poor SNR (SNR<10), a sensitivity and specificity of 98% and 94%,
respectively was achieved with crispness of 99% (Table 2c).
Prognostic Factors:
[0057] With the addition of prognostic criteria to the database two
further classifiers were created, namely, lymph node involvement
and vascular invasion. A small number of known benign or
pre-invasive cases were included in these subsets to assess the
computer's ability to correctly define those cases in which no
nodal involvement or vascular invasion was expected. These benign
or pre-invasive cases were all correctly assigned by the computer
into their respective uninvolved classes.
Lymph Node Involvement:
[0058] There were 31 cases with nodal involvement and 30 without
including 2 DCIS and 3 fibrocystic specimens. All spectra were
included irrespective of SNR. Only those spectra for which complete
pathology and clinical reports were available were included in this
comparison (Table 1). The presence of lymph node metastases was
predicted by SCS with a sensitivity of 96% and specificity of 94%
(Table 3a).
Vascular Invasion:
[0059] SCS-based analysis of spectra was also carried out using
vascular invasion as the criterion. There were 85 spectra for this
analysis (Table 1). A sensitivity of 84% and specificity of 100%
was achieved for the correct determination of vascular invasion,
with an overall accuracy of 92% (Table 3b).
Discussion
[0060] The introduction of preprocessing and SCS analysis of MR
spectra has enhanced the ability to correlate spectroscopic changes
with the pathology of human biopsies. It has also allowed specimens
with sub-optimal cellularity to be analysed, and more importantly,
provided a correlation with clinical criteria not apparent by
visual inspection.
[0061] Visual inspection of spectra, like histopathology, is
limited by the experience and skill of the reader for determining
peak height ratios of metabolites (12). Visual inspection of
spectra and the use of peak height ratio measurements of choline
and creatine discriminated benign from malignant spectra with a
higher degree of accuracy than standard triple assessment of breast
lesions. However, to attain a high degree of accuracy, many spectra
with poor SNR had to be discarded, reducing the effectiveness of
the technique. Previous estimates of cellular material derived from
FNAB; on which to perform MRS analysis reliably, have suggested
that at least 10.sup.6 cells are needed (6).
[0062] By using SCS-derived classifiers it was possible to
distinguish malignant from benign pathologies with higher
sensitivity 92% and specificity 96% for all FNAB spectra including
those with low SNR (Table 2b) than by visual reading of these same
spectra (Table 2a). That SCS-based analysis could more reliably
classify a greater proportion of spectra than could be visually
assessed is testament to the robustness and greater generality of
the computer-based approach.
[0063] The SCS-based result is further improved by presenting to
the computer spectral data with high SNR. The improvement in
sensitivity and specificity gained for spectra with SNR>10
(Table 2c) illustrates this point. Obtaining FNAB with adequate
cell numbers can also enhance the results.
[0064] SCS permits classifiers to recognise patterns containing
more complex information. The classifier has been validated to
diagnose specimens with lymph node involvement and vascular
invasion. The ability of the SCS-derived classifier to predict
lymph node involvement with an accuracy of 95% and vascular
invasion with an accuracy of 92% emphasises the wealth of chemical
information that can be extracted, with the appropriate statistical
approach, from an FNAB of a breast lesion (Table 3).
[0065] A major challenge in breast cancer is the need to identify
and understand the factors that most influence the patient's
prognosis and through timely and appropriate intervention influence
this outcome. Adjuvant therapy can reduce the odds of death during
the first ten years after diagnosis of breast cancer by about of
20-30% (19). The best prognostic indicator of survival in patients
with early breast cancer has been shown to be axillary lymph node
status (20-22).
[0066] Increasingly, sentinel lymph node biopsy is being
investigated as a means to reduce the morbidity and cost of
unnecessary axillary dissection in the two thirds of women with
early invasive breast cancer who prove to be node-negative (23-25),
while preserving the option of full axillary node clearance in
those patients who are node-positive. MRS may possibly determine
nodal involvement from the cellular material derived solely from
the primary tumour, thus limiting the role of sentinel lymph node
biopsy.
[0067] The results, that 52% of patients with lymph node
involvement also had vascular invasion, is in agreement with Barth
et al (26), who showed that peritumoural lympho-vascular invasion
correlated with lymph node involvement (27) and was an independent
predictor of disease free and overall survival (28-31).
[0068] A computer-based statistical classification strategy
providing a robust means of analysing clinical data is becoming a
reality. The power, speed and reproducibility of a computer-based
diagnosis may lead to suitably programmed computers supplanting the
human observer in the clinical laboratory. Patients increasingly
expect certainty in diagnosis and optimum management.
[0069] Several important experimental factors should be noted.
Presently, the MRS method according to the invention has thus far
been demonstrated to work only on aspirated cells from the breast
and not on core biopsies that contain a sufficiently high level of
fat to mask diagnostic and prognostic information. The biopsy
should be representative of the lesion and contain sufficient
cellularity. Furthermore, sample handling is of paramount
importance if the specimen is to be minimally degraded. Quality
control in the spectrometer should be exercised with regard to
pulse sequences, temperature, magnet stability, shimming and water
suppression. The magnetic field at which the database reported
herein was collected is 8.5 Tesla (360 MHz for proton). Because
spectral patterns are frequency dependent, a new classifier should
be developed if one uses different magnetic field strengths.
[0070] The clinical and pathology databases used to train the
classifier should be representative of the full range of
pathologies or the complete demographics of the population, or else
the classifier may be inadequately prepared for all the
possibilities it might encounter in clinical practice. In
developing a database for breast lesions, the training set should
have adequate samples of all the commonly encountered breast
pathologies and be updated upon detection of less common tumour
types.
[0071] The invention is expected to provide a revolutionary impact
on breast cancer management by the use of SCS computerised analysis
of MR spectral features, by obtaining a much higher level of
accuracy in diagnosis of the lesion and also an indication of its
metastatic potential when compared to visual inspection of spectra.
Most importantly, the invention facilitates identification of the
stage of the disease from spectral information of FNAB collected
only from the primary breast lesion.
[0072] The invention allows one to determine pathological
diagnosis, the likelihood of axillary lymph nodal involvement and
tumour vascularisation by SCS-based analysis of proton MR spectra
of a FNAB taken from a primary breast lesion. The SCS-based method
is more accurate and reliable than visual inspection for
identifying complex spectral indicators of diagnosis and
prognosis.
[0073] The ability of an SCS-based analysis of MRS data to provide
prognostic information on lymph node involvement by sampling only
the primary tumour may provide a paradigm shift in the management
of breast cancer. The determination of vascular invasion from the
same cellular material highlights the untapped potential of MRS to
determine prognostic information.
[0074] Although one embodiment of the invention has been shown and
described, numerous variations and modifications will readily occur
to those skilled in the art. The invention is not limited to the
preferred embodiment, and its scope is determined only by the
appended claims.
TABLE-US-00001 TABLE 1 Summary of Clinico-pathological Data Benign/
Lymph Vascular All patients Malignant Nodes Invasion (n = 66) (n =
140) (n = 61) (n = 85) Age Mean .+-. SD 55.8 .+-. 15.4 54.7 .+-. 15
58.4 .+-. 13.2 60.6 .+-. 14.3 (Range) (20-101) (20-90) (29-85)
(29-101) Pathology type Invasive Ductal 89 74 52 68 Invasive
Lobular 8 8 3 5 Mixed Ductal/ 1 1 1 Lob. DCIS 10 1 2* 9*
Fibroadenoma 17 17 Fibrocystic 22 22 3* 2* Papilloma 3 2 Radial
Scar 2 2 Gynaecomastia 1 1 Misc. Benign 13 12 1 Total 166 140 61 85
*These preinvasive and benign lesions were included as known lymph
node negative, vascular invasion negative cases to test the
computer's ability to discern true negatives and positives. They
were all correctly classified by the computer into their respective
classes.
TABLE-US-00002 TABLE 2 Maglignant versus Benign a. Visual
Inspection: Malignant versus Benign Sensitivity Specificity For all
Spectra Malignant (n = 83) vs Benign (n = 57) 85.3% 81.5% Spectra
SNR > 10 Malignant (n = 60) vs Benign (n = 49) 100% 873% b. SCS:
Malignant or Benign (All spectra): (M: 83, B: 57) B M Sensitivity
Specificity PPV % Crisp B 51 4 92.7% 92.4% 92.4% 96.5% M 6 73 92.4%
92.7% 92.7% 95.2% Overall Accuracy: 92.6% Overall % Crisp: 95.7%
(134 of 140) x = 0.922 c. SCS: Malignant or Benign (SNR > 10):
(M: 60, B: 49) B M Sensitivity Specificity PPV % Crisp B 46 3 93.9%
98.3% 98.2% 100.0% M 1 58 98.3% 93.9% 94.1% 98.3% Overall Accuracy:
96.1% Overall % crisp: 99.1% (108 of 109) x = 0.922
TABLE-US-00003 TABLE 3 SCS:-Prognostic Indicators a. Lymph Node
involvement: (P: (Present) 29, A: (Absent) 32) P A Sensitivity
Specificity PPV % Crisp P 25 1 96.2% 93.8% 93.9% 89.7% A 2 30 93.8%
96.2% 96.1% 100% Overall Accuracy: 95.0% Overall % crisp: 95.1% (58
of 61) x = 0.899 b. Vascular Invasion: (P: (Present) 33, A:
(Absent) 52) P A Sensitivity Specificity PPV % Crisp P 26 5 83.9%
100.0% 100.0% 93.9% A 0 49 100.0% 83.9% 86.1% 94.2% Overall
Accuracy: 91.9% Overall % crisp: 94.1% (80 of 85) x = 0.839
REFERENCES
[0075] 1. Kitchen P R, McLennan R, Mursell A. Node-positive breast
cancer: a comparison of clinical and pathological findings and
assessment of axillary clearance. Australian & New Zealand
Journal of Surgery 1980; 50:580-3. [0076] 2. Dixon J M, Senbanjo R
O, Anderson T J, mammography. Risk of breast cancer based on age
and mammographic interpretation [JAMA 1996; 276:39-43]. [0077] 4.
Giard R W, Hermans J. The value of aspiration cytologic examination
of the breast. A statistical review of the medical literature.
Cancer 1992; 69:2104-10. [0078] 5. Delikatny E J, Russell P, Hunter
J C, et al. Proton MR and human cervical neoplasia: ex-vivo
spectroscopy allows distinction of invasive carcinoma of the cervix
from carcinoma in-situ and other preinvasive lesions. Radiology
1993; 188:791-6. [0079] 6. Lean C L, Delbridge L, Russell P, et al.
Diagnosis of follicular thyroid lesions by proton magnetic
resonance on fine needle biopsy. Journal of Clinical Endocrinology
& Metabolism 1995; 80:1306-11. [0080] 7. Lean C L, Newland R C,
Ende D A, Bokey E L, Smith I C, Mountford C E. Assessment of human
colorectal biopsies by 1H MRS: correlation with histopathology.
Magnetic Resonance in Medicine 1993; 30:525-33. [0081] 8. Russell
P, Lean C L, Delbridge L, May G L, Dowd S, Mountford C E. Proton
magnetic resonance and human thyroid neoplasia. I: Discrimination
between benign and malignant neoplasms. American Journal of
Medicine 1994; 96:383-8. [0082] 9. Hahn P, Smith I C, Leboldus L,
Littman C, Somorjai R L, Bezabeh T. The classification of benign
and malignant human prostate tissue by multivariate analysis of 1H
magnetic resonance spectra. Cancer Research 1997; 57:3398-401.
[0083] 10. Rutter A, Hugenholtz H, Saunders J K, Smith I C.
Classification of brain tumors by ex vivo 1H NMR spectroscopy.
Journal of Neurochemistry 1995; 64:1655-61. [0084] 11. Mountford C
E, Lean C L, Hancock R, et al. Magnetic resonance spectroscopy
detects cancer in draining lymph nodes. Invasion & Metastasis
1993; 13:57-71. [0085] 12. Mackinnon W B, Barry P A, Malycha P L,
et al. Fine-needle biopsy specimens of benign breast lesions
distinguished from invasive cancer ex vivo with proton MR
spectroscopy. Radiology 1997; 204:661-6. [0086] 13. Somorjai R L,
Nikulin A E, Pizzi N, et al. Computerized consensus diagnosis: a
classification strategy for the robust analysis of MR spectra. I.
Application to 1H spectra of thyroid neoplasms. Magnetic Resonance
in Medicine 1995; 33:257-63. [0087] 14. Wallace J C, Raaphorst G P,
Somorjai R L, et al. Classification of 1H MR spectra of biopsies
from untreated and recurrent ovarian cancer using linear
discriminant analysis. Magnetic Resonance in Medicine 1997;
38:569-76. [0088] 15. Somorjai R L, Dolenko B, Nikulin A K, et al.
Classification of 1H MR spectra of human brain neoplasms: the
influence of preprocessing and computerized consensus diagnosis on
classification accuracy. Journal of Magnetic Resonance Imaging
1996; 6:437-44. [0089] 16. The Pathology reporting of Breast
Cancer, Recommendations of the Australian Cancer Network. May 1977,
The Stone Press, Epping NSW. [0090] 17. Nikulin A E, Dolenko B,
Bezabeh T, Somorjai R L. Near-optimal region selection for feature
space reduction: novel preprocessing methods for classifying MR
spectra. NMR Biommed 1998; 11:209-217. [0091] 18. Efron B,
Tibshirani R J. An Introduction to the Bootstrap. Monograms of
Statistics and Applied Probability. Vol. 57: Chapman Hill, New
York, 1993. [0092] 19. Weidner N. Prognostic factors in breast
carcinoma. Current Opinion in Obstetrics & Gynecology 1995;
7:4-9. [0093] 20. Fisher B, Ravdin R G, Ausman R K, Slack N H,
Moore G E, Noer R J. Surgical adjuvant chemotherapy in cancer of
the breast: results of a decade of cooperative investigation.
Annals of Surgery 1968; 168:337-56. [0094] 21. Silverstein M J,
Gierson E D, Waisman J R, Senofsky G M, Colburn W J, Gamagami P.
Axillary lymph node dissection for T1a breast carcinoma. Is it
indicated? Cancer 1994; 73:664-7. [0095] 22. Chen Y Y, Schnitt S J.
Prognostic factors for patients with breast cancers cm and smaller.
Breast Cancer Research & Treatment 1998; 51:209-25. [0096] 23.
Cady B. The need to reexamine axillary lymph node dissection in
invasive breast cancer [editorial]. Cancer 1994; 73:505-8. [0097]
24. Beechey-Newman N. Sentinel node biopsy: a revolution in the
surgical management of breast cancer? Cancer Treatment Reviews
1998; 24:185-203. [0098] 25. Hill A D, Mann G B, Borgen P I, Cody H
S, 3rd. Sentinel lymphatic mapping in breast cancer. Journal of the
American College of Surgeons 1999; 188:545-9. [0099] 26. Barth A,
Craig P H, Silverstein M J. Predictors of axillary lymph node
metastases in patients with T1 breast carcinoma [Cancer 1997;
79:1918-22.] [0100] 27. Chadha M, Cabon A B, Friedman P, Vikram B.
PredictorS of axillary lymph node metastases in patients with T1
breast cancer. A multivariate analysis. Cancer 1994; 73:350-3.
[0101] 28. Rosen P P, Groshen S, Saigo P E, Kinne D W, Hellman S.
Pathological prognostic factors in stage I (T1N0M0) and stage II
(T1N1M0) breast carcinoma: a study of 644 patients with median
follow-up of 18 years. Journal of Clinical Oncology 1989;
7:1239-51. [0102] 29. Pinder S E, Ellis I O, Galea M, O'Rouke S,
Blamey R W, Elston C W. Pathological prognostic factors in breast
cancer. III. Vascular invasion: relationship with recurrence and
survival in a large study with long-term follow-up. Histopathology
1994; 24:41-7. [0103] 30. De Laurentiis M, Gallo C, De Placido S,
et al. A predictive index of axillary nodal involvement in operable
breast cancer. British Journal of Cancer 1996; 73:1241-7. [0104]
31. Lauria R, Perrone F, Carlomagno C, et al. The prognostic value
of lymphatic and blood vessel invasion in operable breast cancer.
Cancer 1995; 76:1772-8.
* * * * *