Label-Independent Information Compression for Skin Diseases Recognition

Skin diseases have a widespread impact on people’s lives, making their identification crucial in medical data analysis. Convolutional neural networks (CNNs) are widely used for skin disease recognition, yielding success. However, the current research overlooks the crucial issue of compressing label-independent information. Key information in skin disease images exists in small symptomatic patches, while the rest is extraneous. Regrettably, prevailing CNNs-based methods embed redundancy in skin disease features across convolutional layers, reducing accuracy. This paper introduces an Information Bottleneck theory-based algorithm for selective information propagation. The Hilbert-Schmidt independence criterion (HSIC) is used to calculate the dependency between variables in the algorithm. This algorithm enhances the independence of CNN’s convolutional layers and the input features of the skin disease image during training while improving the dependence of convolutional layers and the label value of the image. This confines superfluous data spread and boosts vital information propagation. Experiments conducted on both a self-collected hypopigmentation dataset and the publicly available ISIC2018 dataset, utilizing ResNet-50, DenseNet-169, Inception-v4, and ConvNeXt-B, confirm the effectiveness of the algorithm. The experimental results demonstrate significant enhancements in accuracy following the application of the algorithm. These four CNNs achieve accuracy improvements of 2.68%, 7.63%, 4.20%, and 3.44% on hypopigmentation dataset, and 0.86%, 0.33%, 2.84%, and 1.19% on ISIC2018 dataset. Thus, this algorithm effectively compresses label-independent information, enhancing skin disease recognition.