FSCL-BC: Federated supervised contrastive learning for breast cancer diagnosis with high sensitivity.
Authors
Affiliations (3)
Affiliations (3)
- Universitat Rovira i Virgili, Department of Computer Engineering and Mathematics, CYBERCAT-Center for Cybersecurity Research of Catalonia, ComSCIAM-Center for Computational Science and Applied Mathematics, Tarragona, Catalonia, Spain; NVISION Systems and Technologies SL, Barcelona, Spain. Electronic address: [email protected].
- Universitat Rovira i Virgili, Department of Computer Engineering and Mathematics, CYBERCAT-Center for Cybersecurity Research of Catalonia, ComSCIAM-Center for Computational Science and Applied Mathematics, Tarragona, Catalonia, Spain.
- NVISION Systems and Technologies SL, Barcelona, Spain.
Abstract
Accurate diagnosis of breast cancer in dense breasts requires expert radiologists to examine multiple ultrasound images per patient. This diagnosis procedure is tedious, time-consuming, and prone to misdiagnosis due to human fatigue. AI-aided diagnosis systems can help alleviate this burden. However, vast amounts of data from multiple hospitals, diverse patient demographics, imaging scanners, and protocols are required to develop accurate, robust, and generalizable AI models. Obtaining such a mixture of data is quite challenging due to privacy concerns, data ownership issues, and regulatory constraints. Moreover, due to subtle visual cues, low resolution, a limited number of labeled samples in hospital datasets, and substantial class imbalance inherent in cancer imaging, deep learning models often overfit to the majority (benign) class. As a result, they struggle to generalize well to unseen data and to achieve high sensitivity, thereby increasing the risk of missed cancer cases. To address these problems, this study aims to develop an accurate AI model for breast cancer prediction from ultrasound images using data from multiple hospitals without requiring data sharing. We introduce FSCL-BC, a privacy-preserving method for the diagnosis of breast cancer from ultrasound images with improved sensitivity, integrating supervised contrastive learning within federated learning. This allows hospitals to keep their data on their premises while collaboratively training the AI model, only exchanging the model parameters trained on their private data. Experimental evaluation within a realistic federated setting shows that FSCL-BC achieved substantially higher diagnostic performance - especially sensitivity - than standardized centralized and vanilla federated training, while providing intrinsic privacy protection. On average, compared to the best baseline, FSCL-BC improved sensitivity by 11.5%, Youden's J index by 7.2%, and F1 score by 3.7%. The improvements in MCC and balanced accuracy were more modest, at 2.8% and 2.4%, respectively. Improved diagnostic accuracy and enhanced patient privacy preservation make FSCL-BC a promising and practical solution for developing AI models for breast cancer diagnosis from ultrasound images, particularly in real-world, resource-constrained settings.