Fed-TransResCaps-HGBOA-GWO-PSO-Net: A Secure and Explainable Federated Hybrid Deep Learning Framework for CT–MRI Medical Image Classification

Abstract

Hospitals rarely share raw imaging data. Privacy statutes, institutional review requirements, and plain commercial caution keep scans behind institutional firewalls, which leaves most large-scale medical imaging models trained on whatever single-site corpus their authors could assemble. Federated learning offers a way out, but it inherits two problems that centralised training never had to face: client data are not identically distributed, and the model updates that cross the network are themselves a disclosure surface. This paper presents Fed-TransResCaps-HGBOA-GWO-PSO-Net, a federated framework for CT–MRI modality classification that addresses both. Three complementary encoders  a frozen ResNet50, a frozen EfficientNetB0, and a lightweight convolutional-transformer branch  are refined by Convolutional Block Attention Modules and concatenated into a 3392-dimensional descriptor. A hybrid Honey-Badger/Grey-Wolf/Particle-Swarm search then prunes this descriptor to a 70% subset, so that only 2374 coordinates reach the classification head and the per-round payload shrinks accordingly. Clients are grouped by label distribution using k-means before aggregation, and updates are combined with FedNova first within and then across clusters, which keeps a pathologically skewed client from dominating the global model. Every accepted update is committed to a SHA-256 hash chain, giving an append-only audit trail of who contributed what and when. Grad-CAM overlays are produced at inference so that a clinician can see which image regions drove a decision. On a 4974-image CT–MRI benchmark partitioned across three deliberately non-IID clients, the framework converged from 56.38% to 98.66% accuracy in twelve communication rounds, with an AUC of 0.9979, specificity of 0.9973, and Matthews correlation of 0.9733. The dominant residual error is a one-sided 2.42% false-negative rate on MRI, which we trace to the label-skewed client partition rather than to the encoder.