S686 Privacy‐Preserving Federated Deep Learning for Robust Multicenter Colorectal Polyp Classification

Introduction: Early colorectal polyp detection via colonoscopy reduces cancer incidence, but privacy regulations hinder pooling endoscopic images. Federated learning (FL) enables collaborative training without centralizing raw data. We created an FL framework using the Endoscopic Image and Video Dataset for Recognition of Colorectal Polyps Morphology and Pathology (ERCPMP) v5 dataset and simulated site‐specific subsets to develop a robust multicenter polyp classifier. Methods: ERCPMPv5 comprises 796 RGB images and 21 videos from 191 patients (Olympus colonoscope, 368 × 256 pixels), annotated with Paris, Pit, JNET, and histopathological (tubular, villous, tubulovillous, hyperplastic, serrated, inflammatory, adenocarcinoma) labels. We randomly split images into 5 nonoverlapping subsets (\∼160/site), ensuring proportional subtype and quality representation. Each site fine‐tuned a ResNet‐50 backbone (pretrained on ImageNet) for 5 local epochs per communication round, using categorical cross‐entropy loss, the Adam optimizer (learning rate 1 × 10⁻⁴), and data augmentation (rotations ±10°, brightness ±15%, flips). Filenames encoded pathology and JNET class for label extraction. After local updates, encrypted weights were federated via Federated Averaging (FedAvg) over 50 rounds. The global model was then fine‐tuned on a validation set (ERCPMP supplementary: n = 120; external: n = 300). Final evaluation used a hold‐out test set (150 images/site; n = 750), balanced by subtype and resolution. Results: The federated model achieved an AUC of 0.95 (95% CI, 0.94–0.96) and 89.2% subtype accuracy. Adenomatous sensitivity was 92.5% (CI, 90.7%–94.2%) and specificity 90.3% (CI, 88.5%–92.0%). Fine‐tuning raised AUC to 0.96 (CI, 0.95–0.97), sensitivity to 94.1%, and specificity to 91.7%. Site‐specific AUCs ranged from 0.94 to 0.96, indicating consistent performance. Conclusion: Our FL framework using ERCPMPv5 achieved performance comparable to centralized training while preserving privacy. Fine‐tuning on supplemental data enhanced generalizability. This approach enables multicenter deployment of AI tools for colorectal polyp detection under strict privacy constraints.