A Multilabel Dermatologic Classification and Interpretability Study for Small Samples and Extremely Unbalanced Classes

Early diagnosis of melanoma is important for patient care, but dermatologists are struggling to keep up with the increasing demand for skin disease care. Computer-aided diagnostic techniques can be used to assist physicians in diagnosis using non-invasive dermoscopy. In this thesis, deep learning techniques are utilized to improve the speed of melanoma diagnosis, alleviate the low accuracy of dermatology datasets due to class imbalance, and combine with a 7-point checklist for melanoma to propose an end-to-end, multi-task, interpretable deep learning diagnostic model. The model consists of three parts: class balancing module, course learning fusion module, and diagnostic decoder, which makes scores for the 7 features of melanoma, and combines the scores and image features to make the final diagnosis, with an average accuracy as high as 74.43%. The experimental results show that the model effectively improves the interpretability of deep learning while ensuring accuracy, and analyzes its advantages in practical applications.