Multiple Instance Learning Using 3D Features for Melanoma Detection

This work presents a contribution to advance current solutions for the problem of melanoma detection based on deep learning (DL) approaches. This is an active research field, which aims to aid on the detection and classification of melanoma (the most lethal type of skin cancer) with non-invasive solutions. By exploiting both 2D and 3D characteristics of the skin lesion surface, the proposed approach advances beyond commonly used colour features of dermoscopic images. Two competing classification methods are exploited, namely Multiple Instance Learning (MIL) and DL, which are combined using an uncertainty-aware decision function. The DL method performs classification resorting to RGB data, while MIL performs 3D feature extraction, selects the most significant set, and performs classification at two different learning instances. The novel aspects of this work include DL uncertainty evaluation mechanisms along with MIL to train a robust ensemble classifier, and also the use of dense light-fields for skin lesion classification. Despite the large class imbalance (often present in medical image datasets), the ensemble model achieves cross-validated melanoma classification accuracy of 84.00% when trained against nevus lesions, and 90.82% accuracy when discriminating against all present lesion types. The results show that, in the absence of discriminative 2D characteristics, the 3D surface provides redeeming results, demonstrating that existing methods can benefit from the proposed method by looking beyond 2D image characteristics.