FedQDFU: Non-IID Federated Quantum Learning for Diabetic Foot Ulcer Classification from Deep Image Features

Abstract

Diabetic foot ulcer (DFU) classification increasingly relies on learning frameworks that exploit medical image representations without centralizing raw data. This study presents FedQDFU, a feature-based federated quantum learning benchmark evaluating a compact variational quantum classifier (VQC) on deep features from VGG16, ResNet50, and DenseNet121, using frozen ImageNet-pretrained backbones as fixed feature extractors over 1020 samples from a single public DFU dataset (696 unique feature vectors after exact-duplicate auditing). Because patient identifiers are unavailable, duplicate-group-aware splitting prevents exact duplicates from crossing the development-test boundary but cannot guarantee patient-level independence. A 6-qubit VQC (43 parameters) is compared with a 129-parameter multilayer perceptron (MLP) under centralized learning, FedAvg, and FedProx across IID and Dirichlet non-IID client distributions (α∈{10,1,0.5,0.1}), with federated-compatible random projection, five-seed evaluation of the primary conditions, and bootstrap confidence intervals on a locked test set. At the primary non-IID setting (α=0.5), the VQC attains a lower mean Matthews correlation coefficient (MCC) than the MLP on all three backbones (e.g., ResNet50: 0.782±0.011 vs. 0.650±0.041 under FedAvg); the paired bootstrap difference is significant only for ResNet50 at the uncorrected pairwise 95% level (ΔMCC=−0.115, 95% CI [−0.218,−0.016]), an effect a Bonferroni correction across backbones is estimated to remove, so this result should be read as suggestive rather than confirmed. No significant centralized-to-FedAvg VQC degradation is detected for any backbone, and a ResNet50 parameter-matched experiment shows the MLP-VQC gap persists even when the MLP is reduced to 41 parameters. Under severe non-IID heterogeneity (α=0.1), the seed-42 partition contains one entirely single-class client (mean pairwise Jensen–Shannon divergence 0.5653 vs. 0.1787 at α=0.5); federated aggregation of this degenerate update yields a fixed-threshold collapse in which all test samples are assigned the positive class (MCC = 0, sensitivity = 1, specificity = 0). This outcome is realization-dependent rather than deterministic: only one of five paired partition/training seeds reproduces the complete collapse.