Enhancing Breast Microcalcification Classification: From Binary to Three-Class Classifier

Abstract

This research explores the optimization of convolutional architectures for breast microcalcification classification and investigates the transition from binary to three-class classifiers with emphasis on the interpretation results of the Grad-CAM method. The study begins by identifying ResNet101 as the most suitable architecture, achieving competitive results across various models. Subsequent experiments reveal the detrimental impact of reducing image size from 674 x 674 to the standard 224 x 224 pixels, attributing decreased model accuracy to the loss of crucial details in already small microcalcifications. Building on these findings, the study introduces a three-class classifier to address limitations observed in binary classification. While the best binary classifier achieves 74,7% accuracy and an MCC of 0,458, interpretation highlights intuitive decision-making based on significant features, albeit with identified shortcomings such as several non-intuitive classification and challenges posed by artifacts and macrocalcifications. Transitioning to a three-class model significantly improves interpretability and model credibility, yielding a 91,7% accuracy and an MCC of 0,767. However, this expansion uncovers new challenges, including misclassification of vascular calcifications and issues with breast implants, emphasizing the complexity of incorporating additional classes.

Keywords

• convolutional neural networks
• artificial intelligence
• mammography
• machine learning
• breast calcifications
• model interpretation