HierCaps-XR: Hierarchical Capsule Networks with Federated Adaptive Routing for Real-Time Pulmonary Nodule Detection in Chest Radiographs

Abstract

 The detection of pulmonary nodules on a chest radiograph continues to be clinically difficult because of the overlap in anatomy, design of lesions, and the space constraint inherent to the traditional convolutional neural network designs. Current deep learning models fail to maintain the geometric relationships between the features identified, resulting in high false positive rates due to rib and vessel interference, which directly affects the reliability of the diagnosis at the population screening level. HierCaps-XR is a Hierarchical Capsule Network with Federated Adaptive Routing, proposed to overcome these constraints by encoding pose vectors with capsules, utilizing attention-gated dynamic routing, and optimizing uncertainty-weighted margin loss using Monte Carlo Dropout. Radiograph-adaptive CLAHE tile normalization, Laplacian-of-Gaussian edge sharpening conditions the input pipeline before hierarchical Laplacian-of-Gaussian edge sharpening. HierCaps-XR was tested on a combined NIH ChestX-ray14 and JSRT corpus of 1,230 test images, with a reported accuracy of 96.8%, recall of 97.2%, specificity of 96.1%, and AUC of 0.982, and end-to-end alert latency of less than two seconds. Attention-gating-based Federated Adaptive Routing has been shown to decrease false positives due to overlapping anatomical structures, making HierCaps-XR an implementation-ready, latency-constrained clinical decision support architecture to support real-time pulmonary screening.