Global and Local Explanations for Skin Cancer Diagnosis Using Prototypes

Providing visual cues to justify the decisions of deep neural networks contributes significantly to increase their explainability. Typical strategies to provide explanations rely on saliency or attention maps that may not be easy to interpret. Moreover, the actual decision-making process is still a black-box. This paper proposes to overcome these limitations using class prototypes, both at the global (image-wide) and local (patch-based) levels. These associate images with the corresponding predictions by measuring similarity with learned image/patch descriptors. Our approach offers both global and local explanations for the decisions of the model, providing a clearer justification that resembles the human reasoning process. The proposed approach was applied to the diagnosis of skin lesions in dermoscopy images, outperforming not only blackbox models, which offer no explanations, but also other state-of-the-art explainable approaches.