SSGNet: semi-supervised multi-path grid network for diagnosing melanoma

Early diagnosis of melanoma can help patients receive timely treatment, thereby increasing the cure rate of the patient. Dermoscopy is an important way for checking melanoma, and in order to precisely diagnose melanoma in dermoscopy images, based on the advantages of multi-scale method and multi-task collaboration in deep learning, a novel semi-supervised multi-path grid network (SSGNet) was constructed. The SSGNet includes a colorization network and a classification network, and these two networks share the encoder part. The classification network can use the colorized network to regularize its own encoder, thereby improving its ability to extract features. In the construction of the SSGNet, firstly, a layer correction block was designed to extract sufficient image features, which used spatial self-attention mechanism and belongs to a residual structure. Secondly, in order to make full use of the multi-scale information of the image, the features of different inputs under the same scale were densely connected to form a novel grid structure. Thirdly, the classification network and the colorization network share the encoder to perform their own tasks. The form of multi-scale input greatly reduces the effect of the lesion’s size in the image on the result, and the grid structure can enable the SSGNet model to make maximum use of features at different scales. We evaluated our proposed SSGNet on the ISIC 2020 dataset, and experiments showed that the SSGNet achieved better results than existing deep learning models.