Federated Learning Architecture for 3D Breast Cancer Image Classification (www.mdpi.com)

Researchers proposed a federated learning (FL) architecture tailored for 3D breast cancer image classification—targeting modalities like Digital Breast Tomosynthesis and 3D MRI—so multiple hospitals can jointly train models without sharing patient data. Their pipeline projects DICOM 3D volumes along the y-axis to extract the most diagnostically relevant, high-contrast slices, converts the task into a standardized 2D representation, and applies intensity normalization (rescaling to [0,1] and dividing by mean total intensity). Local CNNs are trained on each institution’s data (they focused on a binary cancer vs. benign task drawn from a ~5,060-patient dataset) and only model updates (weight deltas/gradients) are sent to a central server for aggregation. Experiments on real-world clinical data show the federated model achieves accuracy comparable to centralized training while preserving privacy. This work is significant because it bridges federated learning and high-dimensional medical imaging: it demonstrates a practical, privacy-preserving route to leverage diverse clinical 3D datasets without contravening regulatory constraints. Key technical implications include an effective preprocessing strategy (3D→2D projection and normalization) that reduces computational burden while retaining salient diagnostic information, and model optimizations that make FL feasible for volumetric imaging. The approach enables broader multi-institution collaboration to improve generalization and early detection, though it also points toward future extensions to handle full 3D architectures and common federated challenges (data heterogeneity, class imbalance).