Discriminative Feature Learning Through Angular Margin-Based Softmax Losses for Breast Cancer Classification

Discriminative Feature Learning Through Angular Margin-Based Softmax Losses for Breast Cancer Classification Pendar Alirezazadeh1, Fadi Dornaika1,2, Abdelmalik Moujahid3 1University of the Basque Country UPV/EHU, Leioa, Spain 2IKERBASQUE, Basque Foundation for Science, Bilbao, Spain 3Universidad Internacional de La Rioja (UNIR), Logro˜no, Spain Published in: Biophysical Reviews and Letters, World Scientific Publishing Company. DOI: 10.1142/S1793048024400022. Accepted July 9, 2024. Abstract When many histopathological breast images with different magnification levels need to be analyzed, diagnosing benign or malignant cancer from the images can be time-consuming. Automatic classification of histopathological images of breast cancer can support the diagnostic workflow in pathology, reducing analysis time. Recently, convolutional neural networks (CNNs) have been used for more accurate classification of breast cancer histopathological images. CNNs typically highlight semantic information to extract discriminative features. However, the traditional softmax loss used by CNNs usually lacks sufficient discriminative power. To address this problem, several angular margin-based softmax loss functions have been proposed, including Large Margin Softmax Loss (A-Softmax), Large Margin Cosine Loss (CosFace), Additive Angular Margin Loss (ArcFace), and Linear ArcFace (Li-ArcFace). All of these improved losses are based on the same concept: maximizing inter-class variance while minimizing intra-class distance. This paper focuses on these four losses and their effectiveness in extracting discriminative features and creating decision margins between classes. Extensive experimental evaluations were conducted on a public and well-known histopathological breast cancer image dataset (BreakHis). Further experiments with the BACH dataset for breast cancer classification and the SARS-CoV-2 CT-scan dataset for COVID-19 detection affirm the generalization capability of angular margin-based softmax losses in medical image classification. Main Contributions •This work investigates and compares four margin-based softmax loss functions for automatic classification of histopathological breast cancer images: A-Softmax, CosFace, ArcFace, and Li-ArcFace. •It demonstrates that these loss functions significantly improve class discrimination (benign vs malignant) and intra-class compactness, outperforming traditional softmax in accuracy and robustness on the BreakHis dataset. 1 •Additional experiments with BACH (breast cancer) and SARS-CoV-2 CT-scan (COVID-19 detection) datasets show that the approach generalizes and provides relevant improvements in various medical classification scenarios. •The authors propose using the lightweight BreastNet architecture as a backbone, which reduces overfitting risk on small medical datasets, and integrates attention mechanisms and multi-scale processing to maximize discriminative feature extraction. •The Li-ArcFace loss, employing a linear logit instead of a cosine, stands out by achieving the best intra-class compactness and inter-class diversity, attaining accuracy rates above 98% at various magnification levels and improving detection in small, imbalanced datasets. •The study highlights that selecting a suitable angular margin is critical for outcomes and suggests future research lines in adaptive margins according to class distribution to reduce bias toward majority classes. Impact of the Paper This study makes important contributions to medical image analysis, specifically the automatic classification of breast cancer using histopathological images. By rigorously evaluating angular margin-based softmax losses, it demonstrates how careful selection of loss functions results in substantial gains in both separability and compactness of learned features. This directly improves the accuracy of computer-aided diagnosis (CAD) systems, leading to more reliable, faster, and potentially more accessible breast cancer screening and diagnostics. Notable areas of impact include: •Improved clinical support: The application of advanced losses to classification improves the trustworthiness and accessibility of AI tools for pathologists, potentially reducing misclassification rates and supporting early intervention. •Transferability: Validation across additional datasets, including COVID-19 CT scans, demonstrates the broad utility of the methods in diverse biomedical imaging scenarios, beyond a single disease domain. •Benchmarking and reproducibility: The paper provides exhaustive experimental results and comparisons, facilitating transparent validation and replication for both academic and clinical research communities. In the ongoing integration of AI into healthcare, this research offers both methodological insights and practical evidence for more precise, generalizable, and explainable image-based diagnosis. Reference Pendar Alirezazadeh, Fadi Dornaika, and Abdelmalik Moujahid. “Discriminative Feature Learning Through Angular Margin-Based Softmax Losses for Breast Cancer Classification.” Biophysical Reviews and Letters, World Scientific Publishing Company, Accepted July 9, 2024. ISSN: 17930480. DOI: 10.1142/S1793048024400022 2