Chaos-Enhanced Prototypical Networks for Few-Shot Medical Image Classification

Abstract

Few-shot learning with Prototypical Networks (ProtoNets) overcomes the challenge of data scarcity in neuroradiology, but conventional ProtoNets are vulnerable to prototype drift and geometric instability attributed to high intra-class variance and magnetic resonance imaging (MRI) artifacts. To tackle this challenge without adding too much complexity to the model, we introduce the Chaos-Enhanced Prototypical Network (CE-ProtoNet) which is composed of a fine-tuned ResNet-18 backbone and a deterministic Logistic Chaos Module (LCM). The LCM injects space-filling, bounded perturbations, along ergodic paths, to mimic continuous intra-class changes during episodic meta-training. Evaluated on a 4-way 5-shot brain tumor MRI benchmark across five random seeds, CE-ProtoNet achieved a peak accuracy of 88.56% ( \(\mu = 86.47\% \pm 1.79\%\) ), a mean specificity of 95.49% (peak 96.19%), and a mean AUC of 0.9696 (peak 0.9746). This greatly improves the baseline ProtoNets ( \(79.34\% \pm 1.02\%\) ) and isotropic noise injection ( \(85.45\% \pm 0.88\%\) ), and outperforms specialized frameworks such as ProtoMed ( \(86.58\% \pm 0.86\%\) ) and matches the performance of the Vision Transformers (ViT-B/16, peak \(88.24\%\) ). CE-ProtoNet is a light weight manifold explorer, providing good diagnostic fidelity at no extra cost of parameters and with minimal latency ( \(< 1.84\%\) ).