Meta-learning for zero-shot skin lesion classification across unseen smartphone devices

Abstract

The biggest challenge to automated classification of skin lesions is robust crossdomain generalization especially when models trained on dermoscopic images are applied to smartphone images. This paper assesses how meta-learning can achieve cross-domain robustness on a strict zero-shot protocol, where no target-domain images, labels, or hyperparameter feedback are used during training or model selection. Models are only trained on HAM10000, BCN20000 dermoscopy datasets and evaluated on PAD-UFES-20 smartphone dataset with a single six-class taxonomy. We compare the baselines of supervised transfer learning with several meta-learning paradigms, which include Prototypical Networks, Meta-Baseline, FEAT, MetaOpt- Net. Experiments are evaluated on convolutional and transformer-based backbones and evaluated in a similar episodic format. Macro-averaged F1-score is used as the main measure of performance because it accounts for class imbalance. Across backbones, the monitored models record a significant drop in source to target performance irrespective of good source domain validation outcomes. Meta-learning techniques mitigate this degradation, with relative macro-F1 gains of about 20–70% more than supervised baselines in a variety of backbone architectures in various configurations. These findings prove that episodic meta-learning can reduce the domain-induced failures in the task of classifying skin lesions in the absence of any target-domain exposure. The suggested evaluation environment captures the actual restrictions of teledermatology implementation and gives empirical data of metalearning as a potential approach towards zero-shot cross-domain medical imaging analysis.