FedPAXA-Net: Prototype-aligned personalized federated learning for multi-client gastrointestinal endoscopy image classification

Federated learning enables collaborative medical-image modelling without centralising raw data, but client heterogeneity and unstable prototype construction can limit personalised learning. This study presents FedPAXA-Net, a prototype-aligned personalised federated learning framework for eight-class gastrointestinal endoscopy image classification across three dataset-derived clients. The final configuration combines an ImageNet-pretrained ConvNeXt-Base backbone, lightweight client-local residual adapters, robust class prototypes estimated using a 10% coordinate-wise trimmed mean, training-only reliability weighting based on class support, within-class dispersion, and temporal stability, and soft-target cosine prototype alignment. Validation-only component analysis supported removal of ProtoAttention and squeeze-and-excitation recalibration from the final architecture. The data partition was additionally audited using exact hashing, repeated source and sequence identifiers, perceptual-hash screening, and manual review, yielding a final partition of 14,909 images comprising 12,509 training, 1200 validation, and 1200 held-out test images. An earlier controlled same-backbone held-out test benchmark showed that Ditto remained the strongest comparative baseline, motivating architectural simplification rather than a superiority claim for the original full configuration. After the revised architecture and audited partition were frozen, five-seed validation achieved a macro-F1 of 0.971079 ± 0.001115 with a 95% Student- t confidence interval of [ 0.969694 , 0.972464 ] . Following completion of validation-based model selection, the five frozen checkpoints were evaluated once on the untouched held-out test set, yielding a macro-F1 of 0.970665 ± 0.000385 with a 95% confidence interval of [ 0.970187 , 0.971142 ] and a mean worst-client F1 of 0.964961. The results support a simpler, auditable prototype-aligned federated design while highlighting the need for matched final-protocol baselines and independent institutional validation.