Breast Cancer Digital Repository Research Articles

Breast Cancer Mass Classification Using Machine Learning, Binary-Coded Genetic Algorithms and an Ensemble of Deep Transfer Learning

Abstract The diagnosis of breast cancer (BC) as early as possible is crucial for increasing the survival rate. Mammography enables finding the breast tissue changes years before they could develop into cancer symptoms. In this study, machine learning methods for BC mass pathology classification have been investigated using the radiologists’ mass annotations on the screen-film mammograms of the Breast Cancer Digital Repository (BCDR). The performances of precomputed features in the BCDR and discrete wavelet transform followed by Radon transform have been investigated by using four sequential feature selections and three genetic algorithms. Feature fusion from craniocaudal and mediolateral oblique views was shown to increase the performance of the classifier. Mass classification has been implemented by deep transfer learning (DTL) using the weights of ResNet50, NASNetLarge and Xception networks. An ensemble of DTL (EDTL) was shown to have higher classification performance than the DTL models. The proposed EDTL has area under the receiver operating curve (AUC) scores of 0.8843 and 0.9089 for mass classification on the region of interest (ROI) and ROI union datasets, respectively. The proposed EDTL has the highest BC mass classification AUC score on the BCDR to date and may be useful for other datasets.

Mass segmentation and classification from film mammograms using cascaded deep transfer learning

Breast cancer is the most common type of cancer among women worldwide. Early breast cancers have a high chance of cure so early diagnosis is critical. Mammography screening allows early detection of breast cancer. There has been an increasing interest in the investigation of computer-aided breast cancer diagnosis recently due in part to the development of the novel high-performing deep learning models. In this study, cascaded deep transfer learning (DTL)-based segmentation methods were investigated to segment mass lesions using mammograms of Breast Cancer Digital Repository. In the first stage, the noise sources in the mammogram background were removed by deep learning-based breast segmentation. In the second stage, the mass segmentation performances of five-layer U-net and U-nets having pre-trained weights from VGG16, ResNet50, and Xception networks in the encoding path were investigated. The performances of attention U-net, residual U-net, MultiResUnet, DeepLabV3Plus, and Unet++ were also investigated. A Unet++ model that uses Xception network weights in the encoder region is proposed. The mass segmentation model predictions were used to estimate mass lesion characterization using DTL. On the test data, an AUC of 0.7829, Dice’s similarity coefficient of 0.6356 and intersection over union of 0.5408 were obtained for mass segmentation using the proposed U-net++Xception model. An AUC of 0.8188 and accuracy of 0.7619 were obtained for mass classification into benign versus malignant. The results show that the proposed DTL pipeline can be used for automatic mass segmentation and classification without using clinical data and may reduce the workload of radiologists.

Breast Cancer Detection in Mammogram Images Using K-Means++ Clustering Based on Cuckoo Search Optimization.

Traditional breast cancer detection algorithms require manual extraction of features from mammogram images and professional medical knowledge. Still, the quality of mammogram images hampers this and extracting high-quality features, which can result in very long processing times. Therefore, this paper proposes a new K-means++ clustering based on Cuckoo Search Optimization (KM++CSO) for breast cancer detection. The pre-processing method is used to improve the proposed KM++CSO method more segmentation efficiently. Furthermore, the interpretability is further enhanced using mathematical morphology and OTSU's threshold. To this end, we tested the effectiveness of the KM++CSO methods on the mammogram image analysis society of the Mini-Mammographic Image Analysis Society (Mini-MIAS), the Digital Database for Screening Mammography (DDSM), and the Breast Cancer Digital Repository (BCDR) dataset through cross-validation. We maximize the accuracy and Jaccard index score, which is a measure that indicates the similarity between detected cancer and their corresponding reference cancer regions. The experimental results showed that the detection method obtained an accuracy of 96.42% (Mini-MIAS), 95.49% (DDSM), and 96.92% (BCDR). On overage, the KM++CSO method obtained 96.27% accuracy for three publicly available datasets. In addition, the detection results provided the 91.05% Jaccard index score.

Improved Threshold Based and Trainable Fully Automated Segmentation for Breast Cancer Boundary and Pectoral Muscle in Mammogram Images

Segmentation of the breast region and pectoral muscle are fundamental subsequent steps in the process of Computer-Aided Diagnosis (CAD) systems. Segmenting the breast region and pectoral muscle are considered a difficult task, particularly in mammogram images because of artefacts, homogeneity among the region of the breast and pectoral muscle, and low contrast along the region of breast boundary, the similarity between the texture of the Region of Interest (ROI), and the unwanted region and irregular ROI. This study aims to propose an improved threshold-based and trainable segmentation model to derive ROI. A hybrid segmentation approach for the boundary of the breast region and pectoral muscle in mammogram images was established based on thresholding and Machine Learning (ML) techniques. For breast boundary estimation, the region of the breast was highlighted by eliminating bands of the wavelet transform. The initial breast boundary was determined through a new thresholding technique. Morphological operations and masking were employed to correct the overestimated boundary by deleting small objects. In the medical imaging field, significant progress to develop effective and accurate ML methods for the segmentation process. In the literature, the imperative role of ML methods in enabling effective and more accurate segmentation method has been highlighted. In this study, an ML technique was built based on the Histogram of Oriented Gradient (HOG) feature with neural network classifiers to determine the region of pectoral muscle and ROI. The proposed segmentation approach was tested by utilizing 322, 200, 100 mammogram images from mammographic image analysis society (mini-MIAS), INbreast, Breast Cancer Digital Repository (BCDR) databases, respectively. The experimental results were compared with manual segmentation based on different texture features. Moreover, evaluation and comparison for the boundary of the breast region and pectoral muscle segmentation have been done separately. The experimental results showed that the boundary of the breast region and the pectoral muscle segmentation approach obtained an accuracy of 98.13% and 98.41% (mini-MIAS), 100%, and 98.01% (INbreast), and 99.8% and 99.5% (BCDR), respectively. On average, the proposed study achieved 99.31% accuracy for the boundary of breast region segmentation and 98.64% accuracy for pectoral muscle segmentation. The overall ROI performance of the proposed method showed improving accuracy after improving the threshold technique for background segmentation and building an ML technique for pectoral muscle segmentation. More so, this article also included the ground-truth as an evaluation of comprehensive similarity. In the clinic, this analysis may be provided as a valuable support for breast cancer identification.

COMPUTER AIDED DIAGNOSIS SYSTEM FOR AUTOMATIC TWO STAGES CLASSIFICATION OF BREAST MASS IN DIGITAL MAMMOGRAM IMAGES

Breast cancer is the most frequent cancer type that is diagnosed in women. The exact causes of such cancer are still unknown. Early and precise detection of breast cancer using mammogram images or biopsy to provide the required medications can increase the healing percentage. There are much current research efforts to developed a computer aided diagnosis (CAD) system based on mammogram images for detecting and classification of breast masses. In this research, a CAD system is developed for automated segmentation and two-stages classification of breast masses. The first stage includes the classification of the masses into seven classes (normal, calcification, circumscribed, spiculated, ill-defined, architectural distortion, asymmetry), which is done using probabilistic neural network (PNN). The second classification stage is to define the severity of abnormality into two classes (Benign and Malignant) which were done using support vector machine (SVM). The results of applying the proposed method on two mammogram image show that the accuracy of detection and segmentation of the breast mass was 99.8% for mammographic image analysis society database (MIAS-DB) with 322 images and 97.5% for breast cancer digital repository (BCDR), BCDR-F03 and BCDR-DN01 with 936 images, while for the first classification stage has accuracy of 97.08%, sensitivity of 98.30% and specificity of 89.8%, and the second classification stage has an accuracy of 99.18%, sensitivity of 98.42% and specificity of 94.90%.

70. Computer aided detection system for the automated classification of clustered microcalcifications in digital mammograms

Purpose Mammography is the most effective and low-cost method for the early detection of breast cancers. The aim of this study was to develop an automated model in order to improve detection and characterization of clustered microcalcifications on full-field digital mammograms (FFDM). Methods We analyzed 98 digital mammograms extracted from the public database Breast Cancer Digital Repository with 276 microcalcification clusters; these images were previously evaluated in double blind by two radiologists which had identified the microcalcification clusters. We identified some regions of interest (ROI) on mammograms after pre-processing by using the Circular Hough Transform algorithm in order to recognize circles corresponding to the microcalcifications. Then, we selected morphological and morphometric features for each ROI, such as Speeded Up Robust Feature, as well as classic statistical features based on Haar Transform. These features were used for giving instructions to a Random Forest classifier in order to recognize – riga clustered microcalcifications normal and abnormal ROIs. Finally, we tested the performance of – riga this automated model in cross-validation and with statistical measurements. Results Statistical results ( Table 1 ) showed good performance of proposed model in automatic detection of breast clustered microcalcifications on FFDM (sensitivity over 90% with 3.54 false positives – riga per image), even in young patients with dense breast tissue (sensitivity over 88%) and according to published data [1] . Also classification performance of clustered microcalcifications (AUC 98,6 ± 0,1, Accuracy 97,6 ± 0,2%, Sensitivity 97,7 ± 0,4%, Specificity 96,2 ± 0,4%) was comparable with the state of the art [2] . Conclusions The proposed automated model enables to improve detection of clustered microcalcifications that could be characterized by difficult diagnostic interpretation especially when occurred in dense breast parenchyma. The proposed model results highly performing and comparable to the state-of-art approaches.

Multi-scale CNN based on region proposals for efficient breast abnormality recognition

Mammographic pattern recognition is one of the most essential tasks in breast cancer diagnosis, and has been studied for several years now to make it suitable and faster. In this paper, we developed a novel deep Convolutional Neural Network (CNN) approach to discriminate normal from abnormal breast tissues using Gaussian pyramid representation for multi-scale analysis (Pyramid-CNN). In order to improve image processing time, we extracted representative region proposals from each mammogram using determinant of the Hessian operator. To improve performance of our model and avoid overfitting, data augmentation techniques based on geometric transformation and sub-histogram equalization were applied on all regions to increase the variance of significant mammographic samples. We evaluated our methodology on the publicly available mammography dataset such as Breast Cancer Digital Repository (BCDR) database. In comparison with the current state-of-the-art methods, the experiments show that our proposed system provides efficient results, achieving the average accuracy of 96.84%, sensitivity of 92.12%, specificity of 98.02%, precision of 92.15%, F1-score of 92.12%, and area under the receiver operating characteristic curve (AUC) of 96.76%. Hence, the study demonstrates that our proposed approach has the potential to significantly improve the conventional recognition and classification strategies for use in advanced clinical application and practice or in general, biomedical imaging field.

Fully automated breast boundary and pectoral muscle segmentation in mammograms.

Breast and pectoral muscle segmentation is an essential pre-processing step for the subsequent processes in computer aided diagnosis (CAD) systems. Estimating the breast and pectoral boundaries is a difficult task especially in mammograms due to artifacts, homogeneity between the pectoral and breast regions, and low contrast along the skin-air boundary. In this paper, a breast boundary and pectoral muscle segmentation method in mammograms is proposed. For breast boundary estimation, we determine the initial breast boundary via thresholding and employ Active Contour Models without edges to search for the actual boundary. A post-processing technique is proposed to correct the overestimated boundary caused by artifacts. The pectoral muscle boundary is estimated using Canny edge detection and a pre-processing technique is proposed to remove noisy edges. Subsequently, we identify five edge features to find the edge that has the highest probability of being the initial pectoral contour and search for the actual boundary via contour growing. The segmentation results for the proposed method are compared with manual segmentations using 322, 208 and 100mammograms from the Mammographic Image Analysis Society (MIAS), INBreast and Breast Cancer Digital Repository (BCDR) databases, respectively. Experimental results show that the breast boundary and pectoral muscle estimation methods achieved dice similarity coefficients of 98.8% and 97.8% (MIAS), 98.9% and 89.6% (INBreast) and 99.2% and 91.9% (BCDR), respectively.

Automatised detection of microcalcification in mammography

Introduction An important area in which an improvement of the imaging techniques would be extremely important, is the diagnosis of breast cancer. For this purpose, mammography is the principal diagnostic tool used. Although it is effective in the early detection of breast cancer, exists a real need for new automatic approaches that can improve the accuracy of detection of breast cancer in mammogram Images. In fact, a computerized system as a second reader can support the radiologist in the interpretation of these exams by reducing the number of false positives and thus, the biopsy procedures not necessary. Purpose In this paper, we propose a Computer Aided Detection System (CAD) for the microcalcification in mammogram images as a diagnostic support tool for radiologists in the analysis. Materials and methods We develop a fully automated tool for (1) pre-processing images using the edge detection process described by Canny which was designed to be an optimal edge detector according to particular criteria; (2) region of Interest extraction; (3) Adapted Hough Transform to identify the microcalcification cluster. The proposed method was evaluated using cases from publicly available mammography dataset such as Breast Cancer Digital Repository (BCDR) database. Results We present the results obtained in terms of accuracy, sensitivity, false positive for image. The proposed system shows results comparable state of the art. Conclusion The proposed method was advantageous in the identification of microcalcifications.

LBP operators on curvelet coefficients as an algorithm to describe texture in breast cancer tissues

In computer-aided diagnosis one of the crucial steps to classify suspicious lesions is the extraction of features. Texture analysis methods have been used in the analysis and interpretation of medical images. In this work we present a method based on the association among curvelet transform, local binary patterns, feature selection by statistical analysis and distinct classification methods, in order to support the development of computer aided diagnosis system. The similar features were removed by the statistical analysis of variance (ANOVA). The understanding of the features was evaluated by applying the decision tree, random forest, support vector machine and polynomial (PL) classifiers, considering the metrics accuracy (AC) and area under the ROC curve (AUC): the rates were calculated on images of breast tissues with different physical properties (commonly observed in clinical practice). The datasets were the Digital Database for Screening Mammography, Breast Cancer Digital Repository and UCSB biosegmentation benchmark. The investigated groups were normal-abnormal and benign-malignant. The association of curvelet transform, local binary pattern and ANOVA with the PL classifier achieved higher AUC and AC values for all cases: the obtained rates were among 91% and 100%. These results are relevant, specially when we consider the difficulties of clinical practice in distinguishing the studied groups. The proposed association is useful as an automated protocol for the diagnosis of breast tissues and may contribute to the diagnosis of breast tissues (mammographic and histopathological images).

Machine learning methods for breast cancer CADx over digital and film mammograms

This work explores the usage of machine learning classifiers (MLCs) to support breast cancer diagnosis over digital and film mammograms. Whichever the source, breast cancer datasets are costly to build, requiring the cooperation of specialists over a tedious process. Often, the choice of digital or film mammograms is limited and we need to understand the implications of using either. Our goal is to use similar data analysis methodology on both kinds of mammograms and understand the behavior of MLC on each one. We trained several MLC configurations on the Breast Cancer Digital Repository, a comprehensive annotated repository of mammograms built in this collaboration and publicly available. We show that intensive use of computer resources provides sound insights on the behaviour of MLC even with small or unbalanced datasets. This supports further decisions on the MLC models generated regarding the need for larger datasets, integration in clinical practice, etc.

An evaluation of image descriptors combined with clinical data for breast cancer diagnosis

Breast cancer computer-aided diagnosis (CADx) may utilize image descriptors, demographics, clinical observations, or a combination. CADx performance was compared for several image features, clinical descriptors (e.g. age and radiologist's observations), and combinations of both kinds of data. A novel descriptor invariant to rotation, histograms of gradient divergence (HGD), was developed to deal with round-shaped objects, such as masses. HGD was compared with conventional CADx features. HGD and 11 conventional image descriptors were evaluated using cases from two publicly available mammography data sets, the digital database for screening mammography (DDSM) and the breast cancer digital repository (BCDR), with 1,762 and 362 instances, respectively. Three experiments were done for each data set according to the type of lesion (i.e., all lesions, masses, and calcifications), resulting in six scenarios. For each scenario, 100 training and test sets were generated via resampling without replacement and five machine learning classifiers were used to assess the diagnostic performance of the descriptors. Clinical descriptors outperformed image descriptors in the DDSM sample (three out of six scenarios), and combining the two kind of descriptors was advantageous in five out of six scenarios. HGD was the best descriptor (or comparable to best) in 8 out of 12 scenarios, demonstrating promising capabilities to describe masses. The combination of clinical data and image descriptors was advantageous in most mammography CADx scenarios. A new descriptor based on the divergence of the gradient (HGD) was demonstrated to be a feasible predictor of breast masses' diagnosis.