MANO-UniDerm: A Lightweight Multimodal Skin Lesion Classification Framework Based on Multi-Scale Attention and Knowledge Distillation

In recent years, deep learning has achieved remarkable progress in skin image classification, especially multimodal methods that combine skin images with patient clinical metadata have demonstrated great potential. However, existing highperformance models generally suffer from large parameter sizes, high computational overhead, and excessive memory utilization. Furthermore, many models only focus on either local details of skin lesions or global contextual information, making it difficult to balance both. Meanwhile, redundant information is easily introduced during cross-modal feature extraction, increasing the inference burden of the model. To address these challenges, we propose a lightweight multimodal classification network named MANO-UniDerm. By introducing a multi-scale attention mechanism and constructing a teacher-student learning framework based on knowledge distillation, our method significantly reduces inference time and memory consumption while improving the classification accuracy of the model. Experimental results show that our proposed method achieves higher classification accuracy with reduced computational overhead in both the ISIC-2019 and PAD-UFES-20 datasets.