Features extraction and medical images description for breast cancer automatic diagnosis

Abstract

Fighting breast cancer remains a major public health concern worldwide, affecting millions of lives each year. Early detection is essential to improve survival rates and treatment results. In recent years, advances in medical imaging, particularly mammography, combined with features-based methods, and deep learning techniques, have significantly improved the accuracy and efficiency of computer-aided diagnosis (CAD) systems. Breast masses and microcalcifications represent the most frequent anomalies with a high risk of malignancy. Most of descriptors found in the literature extract global features and fail to characterize spiculated masses. To address this problem, we focused on developing descriptor adapted to the context of breast cancer, and particularly spiculated masses. PATAR descriptor (Polygon Approximation Triangle-Area Representation) applies a geometric transformation on masses, to simplify the contour while keeping important characteristics like concave and convex spaces. Polygon approximation is done with the Ramer-Douglas-Peucker (RDP) algorithm. After RDP process Triangle-Area Representation (TAR signature) is calculated to quantify and measure spiculations. TAR signature calculates the area made by the corners of polygon. In recent years, deep learning-based models have gained ground in CADx systems. Models like DenseNet, ResNet, and EfficientNet based on Convolutional Neuronal Networks (CNNs) does not perform well facing microcalcifications. Stacking Ensemble learning is a technique that combine multiple model outputs, through meta-learner to make final prediction. We designed an optimal meta-learner composed of fully connected network. Experiment on CBIS- DDSM dataset demonstrate the efficiency of the meta-learner. Boosting is another ensemble learning strategy that learns multiple models sequentially and adjust samples weights after each iteration. In this context, a new boosting algorithm is proposed named Cost-Sensitive Boosting with Error Weighted Adjustments (CSB-EWA). The main contribution in this algorithm consist in using false positive and false negative rates to adjust samples weight to guarantee maximum balance between sensitivity and specificity.