Meta-Transfer Derm-Diagnosis: Exploring Few-Shot Learning and Transfer Learning for Skin Disease Classification in Long-Tail Distribution

Over the past decade, the field of medical image analysis has witnessed remarkable advancements, primarily driven by the development of deep convolutional neural networks and the availability of extensive labeled image datasets. These advancements have notably impacted various tasks, including organ segmentation [1, 2], tumor segmentation [3, 4], and disease detection [5, 6]. Although abundant data exists for common diseases, a significant gap persists in data availability for the over 6,000 known rare diseases, affecting approximately 7% of the global population [7]. The diagnosis of these rare diseases, including some skin conditions, presents unique challenges, particularly due to the limited number of clinical examples available for training deep learning models. The automatic classification of skin lesions exemplifies these challenges, as it is complicated by the long-tailed distribution of skin disease datasets, the subtle variations in lesion appearances, and the overall scarcity of sufficient image data [8]. Various studies have been conducted to address the problem of skin disease classification using deep learning approaches. Recent advancements in this field are mainly in three categories: methods based on transfer learning [12, 13], those relying on few-shot learning [8, 14-19], and approaches using cross-domain few-shot learning [20]. The state of the art models in this domain, such as Meta-DermDiagnosis, MetaMed, and PCN models [8, 14, 18], are designed to extract and learn high-level, domain-specific features during their training process.