Low-Rank Adaptation with Swin Transformers to Enhance Skin Cancer Diagnosis
Keywords: Skin cancer, Swin Transformer, LoRA, Vision Transformer, Medical image classification, HAM10000
TL;DR: We propose a parameter-efficient skin cancer diagnosis model by integrating Low-Rank Adaptation (LoRA) with Swin Transformers, achieving high accuracy while reducing computational cost.
Abstract: Skin cancer is one of the most common forms of cancer worldwide. Automated diagnosis using deep learning has shown promise, but high-performing models like Vision Transformers are often computationally expensive. Swin Transformers are less computationally expensive than ViTs because they use a hierarchical structure with shifted windows for self-attention, limiting computations to local regions instead of the entire image. We propose a Parameter Efficient Fine-tuning (PEFT) method integrating Low-Rank Adaptation (LoRA) into Swin Transformers to reduce model training and inference computational complexity while maintaining high diagnostic performance. Experiments on the standard HAM10000 skin cancer dataset demonstrate the proposed model's effectiveness in skin lesion classification with improved efficiency.
Submission Number: 123
Loading