Classification of mass and normal breast tissue on digital mammograms: multiresolution texture analysis

Methods Used in Computer-Aided Diagnosis for Breast Cancer Detection Using Mammograms: A Review

According to the American Cancer Society's forecasts for 2019, there will be about 268,600 new cases in the United States with invasive breast cancer in women, about 62,930 new noninvasive cases, and about 41,760 death cases from breast cancer. As a result, there is a high demand for breast imaging specialists as indicated in a recent report for the Institute of Medicine and National Research Council. One way to meet this demand is through developing Computer-Aided Diagnosis (CAD) systems for breast cancer detection and diagnosis using mammograms. This study aims to review recent advancements and developments in CAD systems for breast cancer detection and diagnosis using mammograms and to give an overview of the methods used in its steps starting from preprocessing and enhancement step and ending in classification step. The current level of performance for the CAD systems is encouraging but not enough to make CAD systems standalone detection and diagnose clinical systems. Unless the performance of CAD systems enhanced dramatically from its current level by enhancing the existing methods, exploiting new promising methods in pattern recognition like data augmentation in deep learning and exploiting the advances in computational power of computers, CAD systems will continue to be a second opinion clinical procedure.

[Radiomics: Big Data Instead of Biopsies in the Future?]

Precision medicine is increasingly pushed forward, also with respect to upcoming new targeted therapies. Individual characterization of diseases on the basis of biomarkers is a prerequisite for this development. So far, biomarkers are characterized clinically, histologically or on a molecular level. The implementation of broad screening methods (“Omics”) and the analysis of big data – in addition to single markers – allow to define biomarker signatures. Next to “Genomics”, “Proteomics”, and “Metabolicis”, “Radiomics” gained increasing interest during the last years. Based on radiologic imaging, multiple radiomic markers are extracted with the help of specific algorithms. These are correlated with clinical, (immuno-) histopathological, or genomic data. Underlying structural differences are based on the imaging metadata and are often not visible and therefore not detectable without specific software. Radiomics are depicted numerically or by graphs. The fact that radiomic information can be extracted from routinely performed imaging adds a specific appeal to this method. Radiomics could potentially replace biopsies and additional investigations. Alternatively, radiomics could complement other biomarkers and thus lead to a more precise, multimodal prediction. Until now, radiomics are primarily used to investigate solid tumors. Some promising studies in head and neck cancer have already been published.

Radiomics: a new application from established techniques

The increasing use of biomarkers in cancer have led to the concept of personalized medicine for patients. Personalized medicine provides better diagnosis and treatment options available to clinicians. Radiological imaging techniques provide an opportunity to deliver unique data on different types of tissue. However, obtaining useful information from all radiological data is challenging in the era of "big data". Recent advances in computational power and the use of genomics have generated a new area of research termed Radiomics. Radiomics is defined as the high throughput extraction of quantitative imaging features or texture (radiomics) from imaging to decode tissue pathology and creating a high dimensional data set for feature extraction. Radiomic features provide information about the gray-scale patterns, inter-pixel relationships. In addition, shape and spectral properties can be extracted within the same regions of interest on radiological images. Moreover, these features can be further used to develop computational models using advanced machine learning algorithms that may serve as a tool for personalized diagnosis and treatment guidance.

A new and fast image feature selection method for developing an optimal mammographic mass detection scheme

PURPOSE

Selecting optimal features from a large image feature pool remains a major challenge in developing computer-aided detection (CAD) schemes of medical images. The objective of this study is to investigate a new approach to significantly improve efficacy of image feature selection and classifier optimization in developing a CAD scheme of mammographic masses.

METHODS

An image dataset including 1600 regions of interest (ROIs) in which 800 are positive (depicting malignant masses) and 800 are negative (depicting CAD-generated false positive regions) was used in this study. After segmentation of each suspicious lesion by a multilayer topographic region growth algorithm, 271 features were computed in different feature categories including shape, texture, contrast, isodensity, spiculation, local topological features, as well as the features related to the presence and location of fat and calcifications. Besides computing features from the original images, the authors also computed new texture features from the dilated lesion segments. In order to select optimal features from this initial feature pool and build a highly performing classifier, the authors examined and compared four feature selection methods to optimize an artificial neural network (ANN) based classifier, namely: (1) Phased Searching with NEAT in a Time-Scaled Framework, (2) A sequential floating forward selection (SFFS) method, (3) A genetic algorithm (GA), and (4) A sequential forward selection (SFS) method. Performances of the four approaches were assessed using a tenfold cross validation method.

RESULTS

Among these four methods, SFFS has highest efficacy, which takes 3%-5% of computational time as compared to GA approach, and yields the highest performance level with the area under a receiver operating characteristic curve (AUC) = 0.864 ± 0.034. The results also demonstrated that except using GA, including the new texture features computed from the dilated mass segments improved the AUC results of the ANNs optimized using other three feature selection methods. In addition, among 271 features, the shape, local morphological features, fat and calcification based features were the most frequently selected features to build ANNs.

CONCLUSIONS

Although conventional GA is a powerful tool in optimizing classifiers used in CAD schemes of medical images, it is very computationally intensive. This study demonstrated that using a new SFFS based approach enabled to significantly improve efficacy of image feature selection for developing CAD schemes.

Quantifying tumour heterogeneity with CT

Heterogeneity is a key feature of malignancy associated with adverse tumour biology. Quantifying heterogeneity could provide a useful non-invasive imaging biomarker. Heterogeneity on computed tomography (CT) can be quantified using texture analysis which extracts spatial information from CT images (unenhanced, contrast-enhanced and derived images such as CT perfusion) that may not be perceptible to the naked eye. The main components of texture analysis can be categorized into image transformation and quantification. Image transformation filters the conventional image into its basic components (spatial, frequency, etc.) to produce derived subimages. Texture quantification techniques include structural-, model- (fractal dimensions), statistical- and frequency-based methods. The underlying tumour biology that CT texture analysis may reflect includes (but is not limited to) tumour hypoxia and angiogenesis. Emerging studies show that CT texture analysis has the potential to be a useful adjunct in clinical oncologic imaging, providing important information about tumour characterization, prognosis and treatment prediction and response.

A COMPUTER-AIDED SYSTEM FOR MASS DETECTION AND CLASSIFICATION IN DIGITIZED MAMMOGRAMS

This paper presents a computer-assisted diagnostic system for mass detection and classification, which performs mass detection on regions of interest followed by the benign-malignant classification on detected masses. In order for mass detection to be effective, a sequence of preprocessing steps are designed to enhance the intensity of a region of interest, remove the noise effects and locate suspicious masses using five texture features generated from the spatial gray level difference matrix (SGLDM) and fractal dimension. Finally, a probabilistic neural network (PNN) coupled with entropic thresholding techniques is developed for mass extraction. Since the shapes of masses are crucial in classification between benignancy and malignancy, four shape features are further generated and joined with the five features previously used in mass detection to be implemented in another PNN for mass classification. To evaluate our designed system a data set collected in the Taichung Veteran General Hospital, Taiwan, R.O.C. was used for performance evaluation. The results are encouraging and have shown promise of our system.

Heterogeneity of focal breast lesions and surrounding tissue assessed by mammographic texture analysis: preliminary evidence of an association with tumor invasion and estrogen receptor status

AIM

This pilot study investigates whether heterogeneity in focal breast lesions and surrounding tissue assessed on mammography is potentially related to cancer invasion and hormone receptor status.

MATERIALS AND METHODS

Texture analysis (TA) assessed the heterogeneity of focal lesions and their surrounding tissues in digitized mammograms from 11 patients randomly selected from an imaging archive [ductal carcinoma in situ (DCIS) only, n = 4; invasive carcinoma (IC) with DCIS, n = 3; IC only, n = 4]. TA utilized band-pass image filtration to highlight image features at different spatial frequencies (filter values: 1.0-2.5) from fine to coarse texture. The distribution of features in the derived images was quantified using uniformity.

RESULTS

Significant differences in uniformity were observed between patient groups for all filter values. With medium scale filtration (filter value = 1.5) pure DCIS was more uniform (median = 0.281) than either DCIS with IC (median = 0.246, p = 0.0102) or IC (median = 0.249, p = 0.0021). Lesions with high levels of estrogen receptor expression were more uniform, most notably with coarse filtration (filter values 2.0 and 2.5, r(s) = 0.812, p = 0.002). Comparison of uniformity values in focal lesions and surrounding tissue showed significant differences between DCIS with or without IC versus IC (p = 0.0009).

CONCLUSION

This pilot study shows the potential for computer-based assessments of heterogeneity within focal mammographic lesions and surrounding tissue to identify adverse pathological features in mammographic lesions. The technique warrants further investigation as a possible adjunct to existing computer aided diagnosis systems.

Classification of breast tissue in mammograms using efficient coding

Background

Female breast cancer is the major cause of death by cancer in western countries. Efforts in Computer Vision have been made in order to improve the diagnostic accuracy by radiologists. Some methods of lesion diagnosis in mammogram images were developed based in the technique of principal component analysis which has been used in efficient coding of signals and 2D Gabor wavelets used for computer vision applications and modeling biological vision.

Methods

In this work, we present a methodology that uses efficient coding along with linear discriminant analysis to distinguish between mass and non-mass from 5090 region of interest from mammograms.

Results

The results show that the best rates of success reached with Gabor wavelets and principal component analysis were 85.28% and 87.28%, respectively. In comparison, the model of efficient coding presented here reached up to 90.07%.

Conclusions

Altogether, the results presented demonstrate that independent component analysis performed successfully the efficient coding in order to discriminate mass from non-mass tissues. In addition, we have observed that LDA with ICA bases showed high predictive performance for some datasets and thus provide significant support for a more detailed clinical investigation.

Wavelet analysis enables system-independent texture analysis of optical coherence tomography images

Texture analysis for tissue characterization is a current area of optical coherence tomography (OCT) research. We discuss some of the differences between OCT systems and the effects those differences have on the resulting images and subsequent image analysis. In addition, as an example, two algorithms for the automatic recognition of bladder cancer are compared: one that was developed on a single system with no consideration for system differences, and one that was developed to address the issues associated with system differences. The first algorithm had a sensitivity of 73% and specificity of 69% when tested using leave-one-out cross-validation on data taken from a single system. When tested on images from another system with a different central wavelength, however, the method classified all images as cancerous regardless of the true pathology. By contrast, with the use of wavelet analysis and the removal of system-dependent features, the second algorithm reported sensitivity and specificity values of 87 and 58%, respectively, when trained on images taken with one imaging system and tested on images taken with another.

Spiculation-preserving polygonal modeling of contours of breast tumors

Malignant breast tumors typically appear in mammograms with rough, spiculated, or microlobulated contours, whereas most benign masses have smooth, round, oval, or macrolobulated contours. Several studies have shown that shape factors that incorporate differences as above can provide high accuracies in distinguishing between malignant tumors and benign masses based upon their contours only. However, global measures of roughness, such as compactness, are less effective than specially designed features based upon spicularity and concavity. We propose a method to derive polygonal models of contours that preserve spicules and details of diagnostic importance. We show that an index of spiculation derived from the turning functions of the polygonal models obtained by the proposed method yields better classification accuracy than a similar measure derived using a previously published method. The methods were tested with a set of 111 contours of 65 benign masses and 46 malignant tumors. A high classification accuracy of 0.93 in terms of the area under the receiver operating characteristics curve was obtained.

Feature extraction from a signature based on the turning angle function for the classification of breast tumors

Malignant breast tumors and benign masses appear in mammograms with different shape characteristics: the former usually have rough, spiculated, or microlobulated contours, whereas the latter commonly have smooth, round, oval, or macrolobulated contours. Features that characterize shape roughness and complexity can assist in distinguishing between malignant tumors and benign masses. Signatures of contours may be used to analyze their shapes. We propose to use a signature based on the turning angle function of contours of breast masses to derive features that capture the characteristics of shape roughness as described above. We propose methods to derive an index of the presence of convex regions (XR ( TA )), an index of the presence of concave regions (VR ( TA )), an index of convexity (CX ( TA )), and two measures of fractal dimension (FD ( TA ) and FDd ( TA )) from the turning angle function. The methods were tested with a set of 111 contours of 65 benign masses and 46 malignant tumors with different parameters. The best classification accuracies in discriminating between benign masses and malignant tumors, obtained for XR ( TA ), VR ( TA ), CX ( TA ), FD ( TA ), and FDd ( TA ) in terms of the area under the receiver operating characteristics curve, were 0.92, 0.92, 0.93, 0.93, and, 0.92, respectively.

Characterization of mammographic masses using a gradient-based segmentation algorithm and a neural classifier

Computerized methods have recently shown a great potential in providing radiologists with a second opinion about the visual diagnosis of the malignancy of mammographic masses. The computer-aided diagnosis (CAD) system we developed for the mass characterization is mainly based on a segmentation algorithm and on the neural classification of several features computed on the segmented mass. Mass-segmentation plays a key role in most computerized systems. Our technique is a gradient-based one, showing the main characteristic that no free parameters have been evaluated on the data set used in this analysis, thus it can directly be applied to data sets acquired in different conditions without any ad hoc modification. A data set of 226 masses (109 malignant and 117 benign) has been used in this study. The segmentation algorithm works with a comparable efficiency both on malignant and benign masses. Sixteen features based on shape, size and intensity of the segmented masses are extracted and analyzed by a multi-layered perceptron neural network trained with the error back-propagation algorithm. The capability of the system in discriminating malignant from benign masses has been evaluated in terms of the receiver-operating characteristic (ROC) analysis. A feature selection procedure has been carried out on the basis of the feature discriminating power and of the linear correlations interplaying among them. The comparison of the areas under the ROC curves obtained by varying the number of features to be classified has shown that 12 selected features out of the 16 computed ones are powerful enough to achieve the best classifier performances. The radiologist assigned the segmented masses to three different categories: correctly-, acceptably- and non-acceptably-segmented masses. We initially estimated the area under ROC curve only on the first category of segmented masses (the 88.5% of the data set), then extending the classification to the second subclass (reaching the 97.8% of the data set) and finally to the whole data set, obtaining A(z)=0.805+/-0.030, 0.787+/-0.024 and 0.780+/-0.023, respectively.

A ranklet-based image representation for mass classification in digital mammograms

Regions of interest (ROIs) found on breast radiographic images are classified as either tumoral mass or normal tissue by means of a support vector machine classifier. Classification features are the coefficients resulting from the specific image representation used to encode each ROI. Pixel and wavelet image representations have already been discussed in one of our previous works. To investigate the possibility of improving classification performances, a novel nonparametric, orientation-selective, and multiresolution image representation is developed and evaluated, namely a ranklet image representation. A dataset consisting of 1000 ROIs representing biopsy-proven tumoral masses (either benign or malignant) and 5000 ROIs representing normal breast tissue is used. ROIs are extracted from the digital database for screening mammography collected by the University of South Florida. Classification performances are evaluated using the area Az under the receiver operating characteristic curve. By achieving Az values of 0.978 +/- 0.003 and 90% sensitivity with a false positive fraction value of 4.5%, experiments demonstrate classification results higher than those reached by the previous image representations. In particular, the improvement on the Az value over that achieved by the wavelet image representation is statistically relevant with the two-tailed p value <0.0001. Besides, owing to the tolerance that the ranklet image representation reveals to variations in the ROIs' gray-level intensity histogram, this approach discloses to be robust also when tested on radiographic images having gray-level intensity histogram remarkably different from that used for training.

Using Generalized Additive Models for Construction of Nonlinear Classifiers in Computer-Aided Diagnosis Systems

Several investigators have pointed out the possibility of using computer-aided diagnosis (CAD) schemes, as second readers, to help radiologists in the interpretation of images. One of the most important aspects to be considered when the diagnostic imaging systems are analyzed is the evaluation of their diagnostic performance. To perform this task, receiver operating characteristic curves are the method of choice. An important step in nearly all CAD systems is the reduction of false positives, as well as the classification of lesions, using different algorithms, such as neural networks or feature analysis, and several statistical methods. A statistical model more often employed is linear discriminant analysis (LDA). However, LDA implies several limitations in the type of variables that it can analyze. In this work, we have developed a novel approach, based on generalized additive models (GAMs), as an alternative to LDA, which can deal with a broad variety of variables, improving the results produced by using the LDA model. As an application, we have used GAM techniques for reducing the number of false detections in a computerized method to detect clustered microcalcifications, and we have compared this with the results obtained when LDA was applied. Employing LDA, the system achieved a sensitivity of 80.52% at a false-positive rate of 1.90 false detections per image. With the GAM, the sensitivity increased to 83.12% and 1.46 false positives per image.

Texture analysis for tissue discrimination on T1-weighted MR images of the knee joint in a multicenter study: Transferability of texture features and comparison of feature selection methods and classifiers

PURPOSE

To investigate the reproducibility and transferability of texture features between MR centers, and to compare two feature selection methods and two classifiers.

MATERIALS AND METHODS

Coronal T1-weighted MR images of the knees of 63 patients, divided into three groups, were included in the study. MR images were obtained at three different MR centers. Regions of interest (ROIs) were drawn in the bone marrow and fat tissue. Then texture analysis (TA) of the ROIs was performed, and the most discriminant features were identified using Fisher coefficients and POE+ACC (probability of classification error and average correlation coefficients). Based on these features, artificial neural network (ANN) and k-nearest-neighbor (k-NN) classifiers were used for tissue discrimination.

RESULTS

Although the texture features differed among the MR centers, features from one center could be successfully used for tissue discrimination in texture data on MR images from other centers. The best results were achieved using the ANN classifier in combination with features selected by POE+ACC.

CONCLUSION

The differences in texture features extracted from MR images from different centers seem to have only a small impact on the results of tissue discrimination.

Correspondence in texture features between two mammographic views

It is well established that radiologists are better able to interpret mammograms when two mammographic views are available. Consequently, two mammographic projections are standard: mediolateral oblique (MLO) and craniocaudal (CC). Computer-aided diagnosis algorithms have been investigated for assisting in the detection and diagnosis of breast lesions in digitized/digital mammograms. A few previous studies suggest that computer-aided systems may also benefit from combining evidence from the two views. Intuitively, we expect that there would only be value in merging data from two views if they provide complementary information. A measure of the similarity of information is the correlation coefficient between corresponding features from the MLO and CC views. The purpose of this study was to investigate the correspondence in Haralick's texture features between the MLO and CC mammographic views of breast lesions. Features were ranked on the basis of correlation values and the two-view correlation of features for subgroups of data including masses versus calcification and benign versus malignant lesions were compared. All experiments were performed on a subset of mammography cases from the Digital Database for Screening Mammography (DDSM). It was observed that the texture features from the MLO and CC views were less strongly correlated for calcification lesions than for mass lesions. Similarly, texture features from the two views were less strongly correlated for benign lesions than for malignant lesions. These differences were statistically significant. The results suggest that the inclusion of texture features from multiple mammographic views in a CADx algorithm may impact the accuracy of diagnosis of calcification lesions and benign lesions.

Breast cancer diagnosis using three-dimensional ultrasound and pixel relation analysis

Because ultrasound (US) imaging offers benefits compared with other medical imaging techniques, it is used routinely in nearly all hospitals and many clinics. However, the surface features and internal structure of a tumor are not easily demonstrated simultaneously using the traditional 2-D US. The newly developed three-dimensional (3-D) US can capture the morphology of a breast tumor and overcome the limitations of the traditional 2-D US. This study deals with pixel relation analysis techniques for use with 3-D breast US images and compares its performance to 2-D versions of the images. The 3-D US imaging was performed using a Voluson 530 scanner. The rectangular subimages of the volume-of-interest (VOI) were manually selected and the selected VOIs were outlined to include the entire extent of the tumor margin. The databases in this study included 54 malignant and 161 benign tumors. All solid nodules at US belong over C3 (probably benign) according to ACR BI-RADS category. All or some selected 2-D slices were used separately to calculate the diagnosis features for a 3-D US data set. We have proposed and compared several different methods to extract the characteristics of these consecutive 2-D images. As shown in our experiments, the diagnostic results were better than those of the conventional 2-D US. In the experiments, the area index Az under ROC curve of the proposed 3-D US method can achieve 0.9700 +/- 0.0118, but Az of the 2-D US is only 0.8461 +/- 0.0315. The p value of these two Az differences using z test is smaller than 0.01. Furthermore, we can find that the features from only several slices are enough to provide good diagnostic results if the adopted features are modified from the 2-D features.

Textural analysis of contrast-enhanced MR images of the breast

Texture analysis was applied to high-resolution, contrast-enhanced (CE) images of the breast to provide a method of lesion discrimination. Significant differences were seen between benign and malignant lesions for a number of textural features, including entropy and sum entropy. Using logistic regression analysis (LRA), a diagnostic accuracy of A(z) = 0.80 +/- 0.07 was obtained with a model requiring only three parameters. By initially dividing the patient data into training and test datasets, reasonable model robustness was also established. On combining features obtained using textural analysis with lesion size, time to maximum enhancement, and patient age, a diagnostic accuracy of A(z) = 0.92 +/- 0.05 was demonstrated.

Wavelets and imaging informatics: a review of the literature

Modern medicine is a field that has been revolutionized by the emergence of computer and imaging technology. It is increasingly difficult, however, to manage the ever-growing enormous amount of medical imaging information available in digital formats. Numerous techniques have been developed to make the imaging information more easily accessible and to perform analysis automatically. Among these techniques, wavelet transforms have proven prominently useful not only for biomedical imaging but also for signal and image processing in general. Wavelet transforms decompose a signal into frequency bands, the width of which are determined by a dyadic scheme. This particular way of dividing frequency bands matches the statistical properties of most images very well. During the past decade, there has been active research in applying wavelets to various aspects of imaging informatics, including compression, enhancements, analysis, classification, and retrieval. This review represents a survey of the most significant practical and theoretical advances in the field of wavelet-based imaging informatics.

An automatic method to discriminate malignant masses from normal tissue in digital mammograms

Specificity levels of automatic mass detection methods in mammography are generally rather low, because suspicious looking normal tissue is often hard to discriminate from real malignant masses. In this work a number of features were defined that are related to image characteristics that radiologists use to discriminate real lesions from normal tissue. An artificial neural network was used to map the computed features to a measure of suspiciousness for each region that was found suspicious by a mass detection method. Two data sets were used to test the method. The first set of 72 malignant cases (132 films) was a consecutive series taken from the Nijmegen screening programme, 208 normal films were added to improve the estimation of the specificity of the method. The second set was part of the new DDSM data set from the University of South Florida. A total of 193 cases (772 films) with 372 annotated malignancies was used. The measure of suspiciousness that was computed using the image characteristics was successful in discriminating tumours from false positive detections. Approximately 75% of all cancers were detected in at least one view at a specificity level of 0.1 false positive per image.

Applying computer-assisted detection schemes to digitized mammograms after JPEG data compression: an assessment

RATIONALE AND OBJECTIVES

The authors' purpose was to assess the effects of Joint Photographic Experts Group (JPEG) image data compression on the performance of computer-assisted detection (CAD) schemes for the detection of masses and microcalcification clusters on digitized mammograms.

MATERIALS AND METHODS

This study included 952 mammograms that were digitized and compressed with a JPEG-compatible image-compression scheme. A CAD scheme, previously developed in the authors' laboratory and optimized for noncompressed images, was applied to reconstructed images after compression at five levels. The performance was compared with that obtained with the original noncompressed digitized images.

RESULTS

For mass detection, there were no significant differences in performance between noncompressed and compressed images for true-positive regions (P = .25) or false-positive regions (P = .40). In all six modes the scheme identified 80% of masses with less than one false-positive region per image. For the detection of microcalcification clusters, there was significant performance degradation (P < .001) at all compression levels. Detection sensitivity was reduced by 4%-10% as compression ratios increased from 17:1 to 62:1. At the same time, the false-positive detection rate was increased by 91%-140%.

CONCLUSION

The JPEG algorithm did not adversely affect the performance of the CAD scheme for detecting masses, but it did significantly affect the detection of microcalcification clusters.

Quantitative characterization of mass lesions on digitized mammograms for computer-assisted diagnosis

RATIONALE AND OBJECTIVES

To investigate features for discriminating benign from malignant mammographic findings by using computer-aided diagnosis (CAD) and to test the accuracy of CAD interpretations of mass lesions.

METHODS

Fifty-five sequential, mammographically detected mass lesions, referred for biopsy, were digitized for computerized reevaluation with a CAD system. Quantitative features that characterize spiculation were automatically extracted by the CAD system. Data generated by 271 known retrospective cases were used to set reference values indicating the range for malignant and benign lesions. After conventional interpretation of the 55 prospective cases, they were evaluated a second time by the radiologist using the extracted features and the reference ranges. In addition, a pattern-recognition scheme based on the extracted features was used to classify the prospective cases. Accuracy of interpretation with and without the CAD system was evaluated using receiver operating characteristic (ROC) curve analysis.

RESULTS

Sensitivity of the CAD diagnosis for the prospective cases improved from 92% to 100%. Specificity improved significantly from 26.7% to 66.7%. This was accompanied by a significant increase in the accuracy of diagnosis from 56.4% to 81.8% and in the positive predictive value from 51.1% to 71.4%. The Az for the CAD ROC curve significantly increased from 0.73 to 0.90. The performance of the classification scheme was slightly lower than that of the radiologists' interpretation with the CAD system.

CONCLUSIONS

Use of the CAD system significantly improved the accuracy of diagnosis. The findings suggest that the classification scheme may improve the radiologist's ability to differentiate benign from malignant mass lesions in the interpretation of mammograms.

Breast cancer diagnosis using self-organizing map for sonography

The purpose of this study was to evaluate the performance of neural network model self-organizing maps (SOM) in the classification of benign and malignant sonographic breast lesions. A total of 243 breast tumors (82 malignant and 161 benign) were retrospectively evaluated. When a sonogram was performed, the analog video signal was captured to obtain a digitized sonographic image. The physician selected the region of interest in the sonography. An SOM model using 24 autocorrelation texture features classified the tumor as benign or malignant. In the experiment, cases were sampled with k-fold cross-validation (k = 10) to evaluate the performance using receiver operating characteristic (ROC) curves. The ROC area index for the proposed SOM system is 0.9357 +/- 0.0152, the accuracy is 85. 6%, the sensitivity is 97.6%, the specificity is 79.5%, the positive predictive value is 70.8%, and the negative predictive value is 98. 5%. This computer-aided diagnosis system can provide a useful tool and its high negative predictive value could potentially help avert benign biopsies.

Classification of malignant and benign masses based on hybrid ART2LDA approach

A new type of classifier combining an unsupervised and a supervised model was designed and applied to classification of malignant and benign masses on mammograms. The unsupervised model was based on an adaptive resonance theory (ART2) network which clustered the masses into a number of separate classes. The classes were divided into two types: one containing only malignant masses and the other containing a mix of malignant and benign masses. The masses from the malignant classes were classified by ART2. The masses from the mixed classes were input to a supervised linear discriminant classifier (LDA). In this way, some malignant masses were separated and classified by ART2 and the less distinguishable benign and malignant masses were classified by LDA. For the evaluation of classifier performance, 348 regions of interest (ROI's) containing biopsy proven masses (169 benign and 179 malignant) were used. Ten different partitions of training and test groups were randomly generated using an average of 73% of ROI's for training and 27% for testing. Classifier design, including feature selection and weight optimization, was performed with the training group. The test group was kept independent of the training group. The performance of the hybrid classifier was compared to that of an LDA classifier alone and a backpropagation neural network (BPN). Receiver operating characteristics (ROC) analysis was used to evaluate the accuracy of the classifiers. The average area under the ROC curve (A(z)) for the hybrid classifier was 0.81 as compared to 0.78 for the LDA and 0.80 for the BPN. The partial areas above a true positive fraction of 0.9 were 0.34, 0.27 and 0.31 for the hybrid, the LDA and the BPN classifier, respectively. These results indicate that the hybrid classifier is a promising approach for improving the accuracy of classification in CAD applications.

Combined adaptive enhancement and region-growing segmentation of breast masses on digitized mammograms

As an ongoing effort to develop a computer aid for detection of masses on mammograms, we recently designed an object-based region-growing technique to improve mass segmentation. This segmentation method utilizes the density-weighted contrast enhancement (DWCE) filter as a pre-processing step. The DWCE filter adaptively enhances the contrast between the breast structures and the background. Object-based region growing was then applied to each of the identified structures. The region-growing technique uses gray-scale and gradient information to adjust the initial object borders and to reduce merging between adjacent or overlapping structures. Each object is then classified as a breast mass or normal tissue based on extracted morphological and texture features. In this study we evaluated the sensitivity of this combined segmentation scheme and its ability to reduce false positive (FP) detections on a data set of 253 digitized mammograms, each of which contained a biopsy-proven breast mass. It was found that the segmentation scheme detected 98% of the 253 biopsy-proven breast masses in our data set. After final FP reduction, the detection resulted in 4.2 FP per image at a 90% true positive (TP) fraction and 2.0 FPs per image at an 80% TP fraction. The combined DWCE and object-based region growing technique increased the initial detection sensitivity, reduced merging between neighboring structures, and reduced the number of FP detections in our automated breast mass detection scheme.

Design of a high-sensitivity classifier based on a genetic algorithm: application to computer-aided diagnosis

A genetic algorithm (GA) based feature selection method was developed for the design of high-sensitivity classifiers, which were tailored to yield high sensitivity with high specificity. The fitness function of the GA was based on the receiver operating characteristic (ROC) partial area index, which is defined as the average specificity above a given sensitivity threshold. The designed GA evolved towards the selection of feature combinations which yielded high specificity in the high-sensitivity region of the ROC curve, regardless of the performance at low sensitivity. This is a desirable quality of a classifier used for breast lesion characterization, since the focus in breast lesion characterization is to diagnose correctly as many benign lesions as possible without missing malignancies. The high-sensitivity classifier, formulated as the Fisher's linear discriminant using GA-selected feature variables, was employed to classify 255 biopsy-proven mammographic masses as malignant or benign. The mammograms were digitized at a pixel size of 0.1 mm x 0.1 mm, and regions of interest (ROIs) containing the biopsied masses were extracted by an experienced radiologist. A recently developed image transformation technique, referred to as the rubber-band straightening transform, was applied to the ROIs. Texture features extracted from the spatial grey-level dependence and run-length statistics matrices of the transformed ROIs were used to distinguish malignant and benign masses. The classification accuracy of the high-sensitivity classifier was compared with that of linear discriminant analysis with stepwise feature selection (LDAsfs). With proper GA training, the ROC partial area of the high-sensitivity classifier above a true-positive fraction of 0.95 was significantly larger than that of LDAsfs, although the latter provided a higher total area (Az) under the ROC curve. By setting an appropriate decision threshold, the high-sensitivity classifier and LDAsfs correctly identified 61% and 34% of the benign masses respectively without missing any malignant masses. Our results show that the choice of the feature selection technique is important in computer-aided diagnosis, and that the GA may be a useful tool for designing classifiers for lesion characterization.

Voice-activated retrieval of mammography reference images

We undertook this project to integrate context sensitive computer-based educational and decision making aids into the film interpretation and reporting process, and to determine the clinical utility of this method as a guide for further system development. An image database of 347 digital mammography images was assembled and image features were coded. An interface was developed to a computerized speech recognition radiology reporting system which was modified to translate reported findings into database search terms. These observations were used to formulate database search strategies which not only retrieved similar cases from the image database, but also other cases that were related to the index case in different ways. The search results were organized into image sets intended to address common questions that arise during image interpretation. An evaluation of the clinical utility of this method was performed as a guide for further system development. We found that voice dictation of prototypical mammographic cases resulted in automatic retrieval of reference images. The retrieved images were organized into sets matching findings, diagnostic hypotheses, diagnosis, spectrum of findings or diagnoses, closest match to dictated case, or user specified parameters. Two mammographers graded the clinical utility of each form of system output. We concluded that case specific and problem specific image sets may be automatically generated from spoken case dictation. A potentially large number of retrieved images may be divided into subsets which anticipate common clinical problems. This automatic method of context sensitive image retrieval may provide a "continuous" form of education integrated into routine case interpretation.

Computerized characterization of masses on mammograms: the rubber band straightening transform and texture analysis

A new rubber band straightening transform (RBST) is introduced for characterization of mammographic masses as malignant or benign. The RBST transforms a band of pixels surrounding a segmented mass onto the Cartesian plane (the RBST image). The border of a mammographic mass appears approximately as a horizontal line, and possible speculations resemble vertical lines in the RBST image. In this study, the effectiveness of a set of directional textures extracted from the images before the RBST. A database of 168 mammograms containing biopsy-proven malignant and benign breast masses was digitized at a pixel size of 100 microns x 100 microns. Regions of interest (ROIs) containing the biopsied mass were extracted from each mammogram by an experienced radiologist. A clustering algorithm was employed for automated segmentation of each ROI into a mass object and background tissue. Texture features extracted from spatial gray-level dependence matrices and run-length statistics matrices were evaluated for three different regions and representations: (i) the entire ROI; (ii) a band of pixels surrounding the segmented mass object in the ROI; and (iii) the RBST image. Linear discriminant analysis was used for classification, and receiver operating characteristic (ROC) analysis was used to evaluate the classification accuracy. Using the ROC curves as the performance measure, features extracted from the RBST images were found to be significantly more effective than those extracted from the original images. Features extracted from the RBST images yielded an area (Az) of 0.94 under the ROC curve for classification of mammographic masses as malignant and benign.

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.

Computer-aided breast cancer detection and diagnosis of masses using difference of Gaussians and derivative-based feature saliency

A new model-based vision (MBV) algorithm is developed to find regions of interest (ROI's) corresponding to masses in digitized mammograms and to classify the masses as malignant/benign. The MBV algorithm is comprised of five modules to structurally identify suspicious ROI's, eliminate false positives, and classify the remaining as malignant or benign. The focus of attention module uses a difference of Gaussians (DoG) filter to highlight suspicious regions in the mammogram. The index module uses tests to reduce the number of nonmalignant regions from 8.39 to 2.36 per full breast image. Size, shape, contrast, and Laws texture features are used to develop the prediction module's mass models. Derivative-based feature saliency techniques are used to determine the best features for classification. Nine features are chosen to define the malignant/benign models. The feature extraction module obtains these features from all suspicious ROI's. The matching module classifies the regions using a multilayer perceptron neural network architecture to obtain an overall classification accuracy of 100% for the segmented malignant masses with a false-positive rate of 1.8 per full breast image. This system has a sensitivity of 92% for locating malignant ROI's. The database contains 272 images (12 b, 100 microm) with 36 malignant and 53 benign mass images. The results demonstrate that the MBV approach provides a structured order of integrating complex stages into a system for radiologists.

Multifeature analysis of Gd-enhanced MR images of breast lesions

The objective of this study was to determine whether linear discriminant analysis of different independent features of MR images of breast lesions can increase the sensitivity and specificity of this technique. For MR images of 23 benign and 20 malignant breast lesions, three independent classes of features, including characteristics of Gd-DTPA-uptake curve, boundary, and texture were evaluated. The three classes included five, four and eight features each, respectively. Discriminant analysis was applied both within and across the three classes, to find the best combination of features yielding the highest classification accuracy. The highest specificity and sensitivity of the different classes considered independently were as follows: Gd-uptake curves, 83% and 70%; boundary features, 86% and 70%; and texture, 70% and 75%, respectively. A combination of one feature each from the first two classes and age yielded a specificity of 79% and sensitivity of 90%, whereas highest figures of 93% and 95%, respectively, were obtained when a total of 10 features were combined across different classes. Statistical analysis of different independent classes of features in MR images of breast lesions can improve the classification accuracy of this technique significantly.

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.

False-positive reduction technique for detection of masses on digital mammograms: global and local multiresolution texture analysis

We investigated the application of multiresolution global and local texture features to reduce false-positive detection in a computerized mass detection program. One hundred and sixty-eight digitized mammograms were randomly and equally divided into training and test groups. From these mammograms, two datasets were formed. The first dataset (manual) contained four regions of interest (ROIs) selected manually from each of the mammograms. One of the four ROIs contained a biopsy-proven mass and the other three contained normal parenchyma, including dense, mixed dense/fatty, and fatty tissues. The second dataset (hybrid) contained the manually extracted mass ROIs, along with normal tissue ROIs extracted by an automated Density-Weighted Contrast Enhancement (DWCE) algorithm as false-positive detections. A wavelet transform was used to decompose an ROI into several scales. Global texture features were derived from the low-pass coefficients in the wavelet transformed images. Local texture features were calculated from the suspicious object and the peripheral subregions. Linear discriminant models using effective features selected from the global, local, or combined feature spaces were established to maximize the separation between masses and normal tissue. Receiver Operating Characteristic (ROC) analysis was conducted to evaluate the classifier performance. The classification accuracy using global features were comparable to that using local features. With both global and local features, the average area, Az, under the test ROC curve, reached 0.92 for the manual dataset and 0.96 for the hybrid dataset, demonstrating statistically significant improvement over those obtained with global or local features alone. The results indicated the effectiveness of the combined global and local features in the classification of masses and normal tissue for false-positive reduction.

Prescreening entire mammograms for masses with artificial neural networks: preliminary results

RATIONALE AND OBJECTIVES

The authors evaluated the feasibility of combining wavelet transform and artificial neural network (ANN) technologies to prescreen mammograms for masses.

METHODS AND MATERIALS

Fifty-five mammograms (29 with masses and 26 without) were digitized to 100-mm resolution and processed by using wavelet transformation. These wavelets were subjected to a linear output sequential recursive auto-associative memory ANN and cluster analysis with feature vector formation. These vectors were used in two separate experiments-one with 13 cases and another with seven cases held out in a test set-to train feed-forward ANNs to detect the mammograms with a mass. The experiments were repeated with rerandomization of the data, four and six times, respectively.

RESULTS

There was a statistically significant correlation (P < .01) between the network's prediction of a mass and the presence of a mass. With majority voting, the feed-forward ANNs detected masses with 79% sensitivity and 50% specificity.

CONCLUSION

Although preliminary, the combination of wavelet transform and ANN is promising and may provide a viable method to prescreen mammograms for masses with high sensitivity and reasonable specificity.

Computerized classification of malignant and benign microcalcifications on mammograms: texture analysis using an artificial neural network

We investigated the feasibility of using texture features extracted from mammograms to predict whether the presence of microcalcifications is associated with malignant or benign pathology. Eighty-six mammograms from 54 cases (26 benign and 28 malignant) were used as case samples. All lesions had been recommended for surgical biopsy by specialists in breast imaging. A region of interest (ROI) containing the microcalcifications was first corrected for the low-frequency background density variation. Spatial grey level dependence (SGLD) matrices at ten different pixel distances in both the axial and diagonal directions were constructed from the background-corrected ROI. Thirteen texture measures were extracted from each SGLD matrix. Using a stepwise feature selection technique, which maximized the separation of the two class distributions, subsets of texture features were selected from the multi-dimensional feature space. A backpropagation artificial neural network (ANN) classifier was trained and tested with a leave-one-case-out method to recognize the malignant or benign microcalcification clusters. The performance of the ANN was analysed with receiver operating characteristic (ROC) methodology. It was found that a subset of six texture features provided the highest classification accuracy among the feature sets studied. The ANN classifier achieved an area under the ROC curve of 0.88. By setting an appropriate decision threshold, 11 of the 28 benign cases were correctly identified (39% specificity) without missing any malignant cases (100% sensitivity) for patients who had undergone biopsy. This preliminary result indicates that computerized texture analysis can extract mammographic information that is not apparent by visual inspection. The computer-extracted texture information may be used to assist in mammographic interpretation, with the potential to reduce biopsies of benign cases and improve the positive predictive value of mammography.

Adaptive computer-aided diagnosis scheme of digitized mammograms

RATIONALE AND OBJECTIVES

We investigated an adaptive rule-based computer-aided diagnosis (CAD) scheme for digitized mammograms that can be optimized by using an image difficulty index as determined from global measures of image characteristics.

METHODS

First, we defined an image "difficulty" index based on image feature measurements in both the spatial and frequency domains. The CAD scheme then segmented the database into three groups. An image database of 428 digitized mammograms with 220 verified masses was randomly divided into two subsets, one for training (rule-setting) and the other for testing the adaptive CAD scheme. Each of the image difficulty groups in the training set was optimized independently to achieve a low false-positive detection rate while maintaining high detection sensitivity. Scheme performance was then evaluated with the test set, and the results were compared with a global rule-based system that was optimized without the adaptive method.

RESULTS

In this preliminary study, a relatively simple adaptive scheme reduced false-positive mass detections compared with the nonadaptive scheme from 0.85 to 0.53 per image. At the same time sensitivity was not significantly changed.

CONCLUSION

This adaptive CAD scheme has distinct advantages in improving CAD scheme performance as long as the training database includes a large number of cases in each image difficulty group with a variety of true-positive abnormalities.