Dynamic Prototype Memory Network for Robust Multiclass Skin Lesion Diagnosis from Dermoscopic Images

Background: Automated skin lesion diagnosis from dermoscopic images is challenging because lesions of the same disease can vary substantially visually, while visually similar patterns can occur across different disease categories. This article introduces a Dynamic Prototype Memory Network (DPM-Net) for multiclass skin lesion classification. Methods: We first employ a deep visual encoder to extract discriminative feature representations from dermoscopic images. We then compare the extracted lesion embeddings with disease-specific prototypes using a similarity-based prototype assignment mechanism. To enhance the method's representational capability, we dynamically update the prototypes based on the feature distribution of the lesion samples. We incorporate a prototype diversity mechanism to prevent prototype collapse, while a prototype compactness constraint encourages lesion embeddings to remain close to the representative prototypes of their corresponding disease categories. Results: Performance validation shows improved results, with a maximum accuracy of 95.39%. Conclusion: Therefore, the proposed model integrates different components to perform automated multiclass skin lesion diagnosis from dermoscopic images.