A Multi-View Ensemble-Based Weakly Supervised Model for Skin Lesion Images Diagnosis in Dermoscopic Images

Skin cancer has become one of the most common causes of death, and accurate diagnosis of skin lesions is essential for early detection of melanoma. Beyond traditional approaches, computer-aided diagnosis is increasingly applied to cancer detection. A key benefit is that it removes the potential for human error. However, existing methods are unable to achieve quite high accuracy due to noise (such as hair, ink dots, scales, etc.) and the small inter-class and large intra-class differences in skin images. Therefore, a multi-view ensemble-based weakly supervised model for skin lesion image diagnosis in dermoscopy is proposed. In this method, a weakly supervised multi-view(WSM) module is proposed to deal with noise in data images. We propose a multi-scale feature fusion (MFF) module to address the challenge of small inter-class variance and large intra-class variance in skin disease images. The model can integrate features of multiple basic models well, capture information of different scales, and explore and utilize the advantages of different features. We conducted a series of experiments in the open dataset HAM10000. The experimental results show that the performance of the proposed model is superior to that of other models, with an accuracy of 95.90%. In conclusion, our model solves the problems of noise and small inter-class and large intra-class differences in skin images well, and achieves advanced performance in skin disease image recognition.