Promise and Limits of Cross-Architecture Knowledge Distillation for Lightweight Medical Vision Transformers (LiteMed-ViT)

Abstract

Running deep models for medical image analysis at the point of care is challenging with standard architectures: convolutional neural networks (CNNs) are often too heavy for edge devices, while Vision Transformers (ViTs) require massive training datasets to compensate for missing inductive biases. This study investigates how far the dual-token distillation used to train data-efficient ViTs can be pushed when applied across architectures, from a CNN-based teacher into a lightweight ViT student. LiteMed-ViT is introduced as a dual-token ViT featuring a convolutional stem and a class-balanced distillation objective; at 0.234M parameters, LiteMed-ViT carries 47.8x fewer parameters than its ResNet18 teacher at a memory footprint of just 0.89 MB. Evaluated against three compact CNN baselines on three MedMNIST subsets of varying modality and class balance, LiteMed-ViT attains higher accuracy than all three baselines on BloodMNIST (95.05%) and two on PathMNIST (89.15%). Under MedMNIST-C corruptions, it is at least as robust as every compact CNN across all subsets, outperforming them on BloodMNIST and PathMNIST without declining in relative standing as severity increases. Particular attention is devoted to identifying the exact conditions where performance plateaus. On DermaMNIST, LiteMed-ViT's balanced accuracy remains up to 1.9 points below the strongest compact CNNs and does not benefit from higher input resolutions up to 224 pixels, where ResNet18 gains 12.6 points. Stronger teachers do not yield a more accurate student, and overconfident ones lower its DermaMNIST balanced accuracy unless calibrated.