An interpretable CNN-based CAD system for skin lesion diagnosis

Recently, convolutional neural networks have greatly outperformed previous systems based on handcrafted features once the size of public databases has increased. However, these algorithms learn feature representations that are difficult to interpret and analyse. On the other hand, experts require automatic systems to explain their decisions according to clinical criteria which, in the field of melanoma diagnosis, are related to the analysis of dermoscopic features found in the lesions. In recent years, the interpretability of deep networks has been explored using methods that obtain visual features highlighted by neurones or analyse activations to extract more useful information. Following the latter approach, this study proposes a system for melanoma diagnosis that explicitly incorporates dermoscopic feature segmentations into a diagnosis network through a channel modulation scheme. Modulation weights control the influence of the detected visual patterns based on the lesion content. As shown in the experimental section, our design not only improves the system performance on the ISIC 2016 (average AUC of 86.6% vs. 85.8%) and 2017 (average AUC of 94.0% vs. 93.8%) datasets, but also notably enhances the interpretability of the diagnosis, providing useful and intuitive cues to clinicians.