Adaptive computer-aided diagnosis scheme of digitized mammograms

Characterizing anatomical variability in breast CT images

Previous work [Burgess et al., Med. Phys. 28, 419-437 (2001)] has shown that anatomical noise in projection mammography results in a power spectrum well modeled over a range of frequencies by a power law, and the exponent (beta) of this power law plays a critical role in determining the size at which a growing lesion reaches the threshold for detection. In this study, the authors evaluated the power-law model for breast computed tomography (bCT) images, which can be thought of as thin sections through a three-dimensional (3D) volume. Under the assumption of a 3D power law describing the distribution of attenuation coefficients in the breast parenchyma, the authors derived the relationship between the power-law exponents of bCT and projection images and found it to be betasection=betaproj-1. They evaluated this relationship on clinical images by comparing bCT images from a set of 43 patients to Burgess' findings in mammography. They were able to make a direct comparison for 6 of these patients who had both a bCT exam and a digitized film-screen mammogram. They also evaluated segmented bCT images to investigate the extent to which the bCT power-law exponent can be explained by a binary model of attenuation coefficients based on the different attenuation of glandular and adipose tissue. The power-law model was found to be a good fit for bCT data over frequencies from 0.07 to 0.45 cyc/mm, where anatomical variability dominates the spectrum. The average exponent for bCT images was 1.86. This value is close to the theoretical prediction using Burgess' published data for projection mammography and for the limited set of mammography data available from the authors' patient sample. Exponents from the segmented bCT images (average value: 2.06) were systematically slightly higher than bCT images, with substantial correlation between the two (r=0.84).

Automated breast mass detection in 3D reconstructed tomosynthesis volumes: a featureless approach

The purpose of this study was to propose and implement a computer aided detection (CADe) tool for breast tomosynthesis. This task was accomplished in two stages-a highly sensitive mass detector followed by a false positive (FP) reduction stage. Breast tomosynthesis data from 100 human subject cases were used, of which 25 subjects had one or more mass lesions and the rest were normal. For stage 1, filter parameters were optimized via a grid search. The CADe identified suspicious locations were reconstructed to yield 3D CADe volumes of interest. The first stage yielded a maximum sensitivity of 93% with 7.7 FPs/breast volume. Unlike traditional CADe algorithms in which the second stage FP reduction is done via feature extraction and analysis, instead information theory principles were used with mutual information as a similarity metric. Three schemes were proposed, all using leave-one-case-out cross validation sampling. The three schemes, A, B, and C, differed in the composition of their knowledge base of regions of interest (ROIs). Scheme A's knowledge base was comprised of all the mass and FP ROIs generated by the first stage of the algorithm. Scheme B had a knowledge base that contained information from mass ROIs and randomly extracted normal ROIs. Scheme C had information from three sources of information-masses, FPs, and normal ROIs. Also, performance was assessed as a function of the composition of the knowledge base in terms of the number of FP or normal ROIs needed by the system to reach optimal performance. The results indicated that the knowledge base needed no more than 20 times as many FPs and 30 times as many normal ROIs as masses to attain maximal performance. The best overall system performance was 85% sensitivity with 2.4 FPs per breast volume for scheme A, 3.6 FPs per breast volume for scheme B, and 3 FPs per breast volume for scheme C.

Evaluation of computer-aided diagnosis on a large clinical full-field digital mammographic dataset

RATIONALE AND OBJECTIVES

To convert and optimize our previously developed computerized analysis methods for use with images from full-field digital mammography (FFDM) for breast mass classification to aid in the diagnosis of breast cancer.

MATERIALS AND METHODS

An institutional review board approved protocol was obtained, with waiver of consent for retrospective use of mammograms and pathology data. Seven hundred thirty-nine FFDM images, which contained 287 biopsy-proven breast mass lesions, of which 148 lesions were malignant and 139 lesions were benign, were retrospectively collected. Lesion margins were delineated by an expert breast radiologist and were used as the truth for lesion-segmentation evaluation. Our computerized image analysis method consisted of several steps: 1) identified lesions were automatically extracted from the parenchymal background using computerized segmentation methods; 2) a set of image characteristics (mathematic descriptors) were automatically extracted from image data of the lesions and surrounding tissues; and 3) selected features were merged into an estimate of the probability of malignancy using a Bayesian artificial neural network classifier. Performance of the analyses was evaluated at various stages of the conversion using receiver-operating characteristic analysis.

RESULTS

An area under the curve value of 0.81 was obtained in the task of distinguishing between malignant and benign mass lesions in a round-robin by case evaluation on the entire FFDM dataset. We failed to show a statistically significant difference (P = .83) compared to results from our previous study in which the computerized classification was performed on digitized screen-film mammograms.

CONCLUSIONS

Our computerized analysis methods developed on digitized screen-film mammography can be converted for use with FFDM. Results show that the computerized analysis methods for the diagnosis of breast mass lesions on FFDM are promising, and can potentially be used to aid clinicians in the diagnostic interpretation of FFDM.

Incorporation of an iterative, linear segmentation routine into a mammographic mass CAD system

In previous research, we have developed a computer-aided detection (CAD) system designed to detect masses in mammograms. The previous version of our system employed a simple but imprecise method to localize the masses. In this research, we present a more robust segmentation routine for use with mammographic masses. Our hypothesis is that by more accurately describing the morphology of the masses, we can improve the CAD system's ability to distinguish masses from other mammographic structures. To test this hypothesis, we incorporated the new segmentation routine into our CAD system and examined the change in performance. The developed iterative, linear segmentation routine is a gray level-based procedure. Using the identified regions from the previous CAD system as the initial seeds, the new segmentation algorithm refines the suspicious mass borders by making estimates of the interior and exterior pixels. These estimates are then passed to a linear discriminant, which determines the optimal threshold between the interior and exterior pixels. After applying the threshold and identifying the object's outline, two constraints on the border are applied to reduce the influence of background noise. After the border is constrained, the process repeats until a stopping criterion is reached. The segmentation routine was tested on a study database of 183 mammographic images extracted from the Digital Database for Screening Mammography. Eighty-three of the images contained 50 malignant and 50 benign masses; 100 images contained no masses. The previously developed CAD system was used to locate a set of suspicious regions of interest (ROIs) within the images. To assess the performance of the segmentation algorithm, a set of 20 features was measured from the suspicious regions before and after the application of the developed segmentation routine. Receiver operating characteristic (ROC) analysis was employed on the ROIs to examine the discriminatory capabilities of each individual feature before and after the segmentation routine. A statistically significant performance increase was found in many of the individual features, particularly those describing the mass borders. To examine how the incorporation of the segmentation routine affected the performance of the overall CAD system, free-response ROC (FROC) analysis was employed. When considering only malignant masses, the FROC performance of the system with the segmentation routine appeared better than the previous system. When detecting 90% of the malignant masses, the previous system achieved 4.9 false positives per image (FPpI) compared to the post-segmentation system's 4.2 FPpI. At 80% sensitivity, the respective FPpI were 3.5 and 1.6.

Computer-Aided Detection Schemes: The Effect of Limiting the Number of Cued Regions in Each Case

OBJECTIVE

We assessed performance changes of a mammographic computer-aided detection scheme when we restricted the maximum number of regions that could be identified (cued) as showing positive findings in each case.

MATERIALS AND METHODS

A computer-aided detection scheme was applied to 500 cases (or 2,000 images), including 300 cases in which mammograms showed verified malignant masses. We evaluated the overall case-based performance of the scheme using a free-response receiver operating characteristic approach, and we measured detection sensitivity at a fixed false-positive detection rate of 0.4 per image after gradually reducing the maximum number of cued regions allowed for each case from seven to one.

RESULTS

The original computer-aided detection scheme achieved a maximum case-based sensitivity of 97% at 3.3 false-positive detected regions per image. For a detection decision score set at 0.565, the scheme had a 79% (237/300) case-based sensitivity, with 0.4 false-positive detected regions per image. After limiting the number of maximum allowed cued regions per case, the false-positive rates decreased faster than the true-positive rates. At a maximum of two cued regions per case, the false-positive rate decreased from 0.4 to 0.21 per image, whereas detection sensitivity decreased from 237 to 220 masses. To maintain sensitivity at 79%, we reduced the detection decision score to as low as 0.36, which resulted in a reduction of false-positive detected regions from 0.4 to 0.3 per image and a reduction in region-based sensitivity from 66.1% to 61.4%.

CONCLUSION

Limiting the maximum number of cued regions per case can improve the overall case-based performance of computer-aided detection schemes in mammography.

Mass discrimination in mammography

RATIONALE AND OBJECTIVES

In a previous publication concerning detection of masses in mammograms it was shown that the amplitude (contrast) required for detection increased as mass size increased. The work presented here was designed to measure the variation of amplitude threshold for discrimination between masses as a function of lesion size.

MATERIALS AND METHODS

A hybrid image method with digitized masses added to digitized normal mammograms was used. The masses were extracted from surgical specimen radiographs. Observer experiments were performed using the two-alternative forced-choice method with images displayed on a computer monitor. There were two tasks: (1) discrimination between a ductal carcinoma and a fibroadenoma, and (2) discrimination between two ductal carcinomas. Masses were scaled to cover the linear size range from 1 to 16 mm. Three observers took part, two physicists and a radiologist.

RESULTS

The discrimination contrast-detail (CD) diagrams were found to have minimum threshold amplitudes at lesion sizes near 4 mm. The detection results had demonstrated an unusual contrast-detail diagram form with threshold amplitudes monotonically increasing with lesion size for lesions larger than 1 mm, which was opposite the usual result for image noise. Discrimination thresholds or masses larger than 4 mm were approximately 1.5-2 times those reported for detection of the lesions.

CONCLUSION

The detection results had been explained using a relatively simple model based on signal detection theory with some characteristics of the human visual system included. The observer model cannot explain the discrimination results, so additional complexity must be introduced to the observer model.

Self-organizing map for cluster analysis of a breast cancer database

The purpose of this study was to identify and characterize clusters in a heterogeneous breast cancer computer-aided diagnosis database. Identification of subgroups within the database could help elucidate clinical trends and facilitate future model building. A self-organizing map (SOM) was used to identify clusters in a large (2258 cases), heterogeneous computer-aided diagnosis database based on mammographic findings (BI-RADS) and patient age. The resulting clusters were then characterized by their prototypes determined using a constraint satisfaction neural network (CSNN). The clusters showed logical separation of clinical subtypes such as architectural distortions, masses, and calcifications. Moreover, the broad categories of masses and calcifications were stratified into several clusters (seven for masses and three for calcifications). The percent of the cases that were malignant was notably different among the clusters (ranging from 6 to 83%). A feed-forward back-propagation artificial neural network (BP-ANN) was used to identify likely benign lesions that may be candidates for follow up rather than biopsy. The performance of the BP-ANN varied considerably across the clusters identified by the SOM. In particular, a cluster (#6) of mass cases (6% malignant) was identified that accounted for 79% of the recommendations for follow up that would have been made by the BP-ANN. A classification rule based on the profile of cluster #6 performed comparably to the BP-ANN, providing approximately 25% specificity at 98% sensitivity. This performance was demonstrated to generalize to a large (2177) set of cases held-out for model validation.

Computerized assessment of tissue composition on digitized mammograms

RATIONALE AND OBJECTIVES

The authors developed a computerized method for the quantitative assessment of breast tissue composition on digitized mammograms.

MATERIALS AND METHODS

Three radiologists were asked to review 200 digitized mammograms and independently provide a Breast Imaging Reporting and Data System-like rating for breast tissue composition on a scale of 0 to 4. These values were incorporated into a "consensus" rating that was used as a reference point in the development and evaluation of a computerized method. After tissue segmentation that excluded nontissue areas, a set of quantitative features was computed. A computerized summary index that attempts to reproduce the radiologists' ratings was developed. Correlation coefficients (Pearson r) were used to compare the computerized index with the consensus ratings.

RESULTS

Some individual features computed for the relatively dense breast areas showed good correlation (r > 0.8) with the radiologists' subjective ratings. The summary index of tissue composition demonstrated a significant correlation (r = 0.87), as well.

CONCLUSION

Computerized methods that show good correlation with radiologists' ratings of breast tissue composition can be developed.

Performance gain in computer-assisted detection schemes by averaging scores generated from artificial neural networks with adaptive filtering

The authors investigated a new method to optimize artificial neural networks (ANNs) with adaptive filtering used in computer-assisted detection schemes in digitized mammograms and to assess performance changes when averaging classification scores from three sets of optimized schemes. Two independent training and testing image databases involving 978 and 830 digitized mammograms, respectively, were used in this study. In the training data set, initial filtering and subtraction resulted in the identification of 592 mass regions and 3790 suspicious, but actually negative regions. These regions (including both true-positive and negative regions) were segmented into three subsets three times based on the calculation of the values of three features as segmentation indices. The indices were "mass" size multiplied by their digital value contrast, conspicuity, and circularity. Nine ANN-based classifiers were separately optimized using a genetic algorithm for each subset of regions. Each region was assigned three classification scores after applying the three adaptive ANNs. The performance gain of the CAD scheme after averaging the three scores for each suspicious region was tested using an independent data set and a ROC methodology. The experimental results showed that the areas under ROC curves (Az) for the testing database using three sets of optimized ANNs individually were 0.84+/-0.01, 0.83+/-0.01, and 0.84+/-0.01, respectively. The between-index correlations of three A values were 0.013, -0.007, and 0.086. Similar to averaging diagnostic ratings from independent observers, by averaging three ANN-generated scores for each testing region, the performance of the CAD scheme was significantly improved (p<0.001) with Az value of 0.95+/-0.01.

Human observer detection experiments with mammograms and power-law noise

We determined contrast thresholds for lesion detection as a function of lesion size in both mammograms and filtered noise backgrounds with the same average power spectrum, P(f)=B/f3. Experiments were done using hybrid images with digital images of tumors added to digitized normal backgrounds, displayed on a monochrome monitor. Four tumors were extracted from digitized specimen radiographs. The lesion sizes were varied by digital rescaling to cover the range from 0.5 to 16 mm. Amplitudes were varied to determine the value required for 92% correct detection in two-alternative forced-choice (2AFC) and 90% for search experiments. Three observers participated, two physicists and a radiologist. The 2AFC mammographic results demonstrated a novel contrast-detail (CD) diagram with threshold amplitudes that increased steadily (with slope of 0.3) with increasing size for lesions larger than 1 mm. The slopes for prewhitening model observers were about 0.4. Human efficiency relative to these models was as high as 90%. The CD diagram slopes for the 2AFC experiments with filtered noise were 0.44 for humans and 0.5 for models. Human efficiency relative to the ideal observer was about 40%. The difference in efficiencies for the two types of backgrounds indicates that breast structure cannot be considered to be pure random noise for 2AFC experiments. Instead, 2AFC human detection with mammographic backgrounds is limited by a combination of noise and deterministic masking effects. The search experiments also gave thresholds that increased with lesion size. However, there was no difference in human results for mammographic and filtered noise backgrounds, suggesting that breast structure can be considered to be pure random noise for this task. Our conclusion is that, in spite of the fact that mammographic backgrounds have nonstationary statistics, models based on statistical decision theory can still be applied successfully to estimate human performance.

Computerized localization of breast lesions from two views. An experimental comparison of two methods

RATIONALE AND OBJECTIVES

The authors compared two computerized methods, the arc and cartesian straight-line, for the localization of breast lesions in two mammographic views.

METHODS

A total of 571 craniocaudal and 571 mediolateral oblique matched mammographic image pairs (or 1142 individual images) depicting 290 pathology-verified masses on both views were selected from our image database. Using a previously developed computer-aided detection scheme, all 290 masses and 3992 suspicious but negative regions were identified. After pairing all identified regions from both views, all masses (true-positive-true-positive matched pairs) and a total of 10330 false-positive pairs (including false-positive-false-positive, true-positive-false-positive, and false-positive-true positive pairs) were assessed as to their position in relation to the nipple using both the arc and the cartesian straight-line methods. Receiver operating characteristic methodology was used to evaluate the performance levels for each method in determining, based solely on location, whether a pair of suspicious regions represented a true mass or a false-positive combination.

RESULTS

The areas under the receiver operating characteristic curves (Az) were 0.79 and 0.78 for the arc and cartesian straight-line methods, respectively. The difference between the two techniques (as measured by Az) was not statistically significant (P > 0.99).

CONCLUSIONS

These preliminary results demonstrated that the two methods are comparable in identifying true masses from triangulated observations on two views. However, the arc method is somewhat favorable because only the nipple location is required for localization.

Feature selection for computerized mass detection in digitized mammograms by using a genetic algorithm

RATIONALE AND OBJECTIVES

To investigate optimization of feature selection for computerized mass detection in digitized mammograms, and to compare the effectiveness of a genetic algorithm (GA) in such optimization with that of an "exhaustive" search of all feature permutations.

MATERIALS AND METHODS

A Bayesian belief network (BBN) was used to classify positive and negative regions for masses depicted in digitized mammograms; 20 features were computed for each of 592 positive and 3,790 negative regions in two databases. Conditional probabilities for the BBN were computed by using a "training" database of 288 positive and 2,204 negative regions. Performance was measured by the area under the receiver operating characteristic curve (A) by using the remainder database (304 positive and 1,586 negative regions). The optimal set was first found by using an "exhaustive" (complete permutation) searching method. A GA-based search for the optimal set then was applied, and the results of the two approaches were compared.

RESULTS

As the number of features in the classifier increased, the A value increased until it reached a maximum performance for 11 features of 0.876 +/- 0.008. The A value then decreased monotonically as the number of features increased from 11 to 20. Using 100 random chromosomes (seeds) in the first generation, the GA identified the same optimal set of features but reduced the total computation time by a factor of 65.

CONCLUSION

A GA-based search might be an efficient and effective approach to selecting an optimal feature set.

Visual signal detection with two-component noise: low-pass spectrum effects

Detection of signals in natural images and scenes is limited by both noise and structure. The purpose of this study is to investigate phenomenological issues of signal detection in two-component noise. One component had a broadband (white) spectrum designed to simulate image noise. The other component was filtered to simulate two classes of low-pass background structure spectra: Gaussian-filtered noise and power-law noise. Measurements of human and model observer performance are reported for several aperiodic signals and both classes of background spectra. Human results are compared with two classes of observer models and are fitted very well by suboptimal prewhitening matched filter models. The nonprewhitening model with an eye filter does not agree with human results when background-noise-component power spectrum bandwidths are less than signal energy bandwidths.

Incorporation of a set enumeration trees-based classifier into a hybrid computer-assisted diagnosis scheme for mass detection

RATIONALE AND OBJECTIVES

The authors evaluated whether a hybrid classifier of two independent computer-aided diagnosis (CAD) schemes, the set enumeration (SE) trees approach and an artificial neural network (ANN), could improve the detection of masses on digitized mammograms. The potential benefits resulting from the interpretability of the SE trees model was also explored.

MATERIALS AND METHODS

Two hundred thirty verified mass regions and 230 negative but suspicious regions were randomly selected from 618 digitized mammograms. Each region was represented by a 24-parameter feature vector. These features were used as input data for the SE trees and ANN-based schemes. After the positive and negative regions were randomly segmented into five exclusive partitions, a fivefold cross-validation method was applied to evaluate and compare the performance of the SE trees, ANN, and hybrid system in the identification of masses.

RESULTS

The performance of the SE trees approach was comparable to that of the ANN. The average area under the receiver operating characteristic (ROC) curves for all five partitions was 0.88 (standard deviation, 0.04). Owing to the relatively low correlation between the region-based results of the SE trees and ANN methods, the hybrid classifier yielded a significantly improved performance, with an area under the ROC curve of 0.94 (standard deviation, 0.02; P < .05).

CONCLUSION

The hybrid CAD scheme significantly improved performance. The amenability of the SE trees models to interpretation may aid in the assessment of the importance of specific features.

Adequacy testing of training set sample sizes in the development of a computer-assisted diagnosis scheme

RATIONALE AND OBJECTIVES

The authors assessed the performance changes of a computer-assisted diagnosis (CAD) scheme as a function of the number of regions used for training (rule-setting).

MATERIALS AND METHODS

One hundred twenty regions depicting actual masses and 400 suspicious but actually negative regions were selected as a testing data set from a database of 2,146 regions identified as suspicious on 618 mammograms. An artificial neural network using 24 and 16 region-based features as input neurons was applied to classify the regions as positive or negative for the presence of a mass. CAD scheme performance was evaluated on the testing data set as the number of regions used for training increased from 60 to 496.

RESULTS

As the number of regions in the training sets increased, the results decreased and plateaued beyond a sample size of approximately 200 regions. Performance with the testing data set continued to improve as the training data set increased in size.

CONCLUSION

A trend in a system's performance as a function of training set size can be used to assess adequacy of the training data set in the development of a CAD scheme.