Multi-Scale DCNN with Dynamic Weight and Part Cross-Entropy Loss for Skin Lesion Diagnosis

Accurately diagnosing skin lesion disease is a challenging task. Although present methods often use the multi-branch structure to get more clues, the rigescent methods of cropping zone and fusing branch results fail to handle the instability of the disease zone and the difference in branch results, which leads to improper cropping and degrades Deep Convolutional Neural Networks (DCNN)'s performance. To address these problems, we propose a Multi-scale DCNN with Dynamic weight and Part cross-entropy loss model (namely MDP-DCNN) to bootstrap skin lesion diagnosis. Inspired by the object detection method, the multi-scale structure adjusts the cropping position based on the Gradient-weighted Class Activation Mapping (Grad-CAM) center. It enables the model to adapt to the disease zone variety in position and size. The dynamic weight structure alleviates the negative influence of branch differences by comparing the grey-cropped zone and its CAM. Moreover, we also propose the part cross-entropy loss to deal with the over-fitting problem. This optimizes the non-targeted label to decrease the influence on other labels' stability when the prediction is wrong. We conduct our model on the ISIC-2017 and ISIC-2018 datasets. Experiments demonstrate that MDP-DCNN achieves excellent results in skin lesion classification without external data. Multi-scale DCNN with dynamic weight and part loss function verifies its advantages in enhancing diagnosis accuracy.