Meta-Learning Coupled with Transfer Learning for Improved Few-Shot Classification of Cardiac MR Images.

In this paper, we propose a novel meta-transfer learning framework for few-shot classification of cardiac magnetic resonance (CMR) images, addressing the challenge of limited annotated medical data, that affects the effectiveness of traditional deep learning methods. Our method combines pre-trained deep neural networks to perform both binary (healthy vs diseased) and 5-class: normal cardiac function (NOR), dilated cardiomyopathy (DCM), hypertrophic cardiomyopathy (HCM), heart failure due to myocardial infractions (MINF) and abnormal RV (ARV) classification, on top of which different meta-learning few shot classifications were added. We investigated both experiments with same and different classes in the pre-training and meta-training phases. By using few-shot learning mechanisms across 1-shot, 3-shot and 5-shot scenarios, the model demonstrates significant improvements in classification accuracy and generalization, even with minimal number of labeled samples. Achieved accuracy demonstrates competitive results with state-of-the-art methods with studies on the ACDC dataset, highlighting the potential of meta-transfer learning to improve diagnostic workflows in cardiac imaging.Clinical Relevance-Proposed method contributes to diagnostic workflows and treatment planning for cardiac conditions, particularly in clinical environment, where acquiring large, annotated datasets is challenging.