DermaKNet is a Computer Aided Diagnosis (CAD) system that incorporates the expert knowledge from dermatologists into the paradigm of Convolutional Neural Networks (CNN). The broad adoption of deep learning techniques in the task of automatic analysis of skin lesions had a great impact on the system performance at the expense of a considerable loss of interpretability compared to traditional systems using hand-crafted features.
Hence, aiming to get the best of two worlds, we designed novel blocks that model aspects of the lesions that are of special relevance for clinicians in their diagnosis, and incorporated them to CNN architectures addressing lesion segmentation, dermoscopic feature segmentation and diagnosis.
Our experiments in the dataset of the ISIC 2017 Challenge “Skin Lesion Analysis Toward Melanoma Detection” demonstrate that the incorporation of these subsystems not only improves the performance, but also enhances the diagnosis by providing more interpretable outputs. These outputs can be easily aligned with aspects considered by dermatologists in their diagnosis, such as lesion asymmetry or the presence of relevant dermoscopic features and their location within a lesion. In consequence, DermaKNet might be used no only as a pure CAD system that helps clinicians in their daily practice, but also as a supporting tool to train novel dermatologists.
Code
https://github.com/igondia/matconvnet-dermoscopy
References
I. Gonzalez Diaz, “DermaKNet: Incorporating the knowledge of dermatologists to Convolutional Neural Networks for skin lesion diagnosis,” in IEEE Journal of Biomedical and Health Informatics, vol. PP, no. 99, pp. 1-1. DOI: 10.1109/JBHI.2018.2806962