Multi-view compression and collaboration for skin disease diagnosis

In the field of skin disease diagnosis based on Convolutional Neural Networks (CNNs), there are currently two challenges. Firstly, there is a significant amount of label-independent information present in skin disease images. This information significantly affects the CNN’s ability to recognize skin disease. Finding an effective way to remove this label-independence is a challenging problem. Secondly, most research focuses solely on information-limited RGB images. It is imperative to introduce additional color space views. Hence, there is a need to investigate which combinations of views are most effective for skin disease diagnosis. To address these two issues, this study first employs the information bottleneck theory to guide convolution operations, retaining relevant skin lesion information while filtering out irrelevant details. Secondly, through a view selection method, a combination of RGB, HSL, and YCbCr was chosen from seven views, which exhibited the best performance. A multi-view compression and collaboration (MCC) framework was constructed based on these two approaches. MCC assists CNNs in removing label-independent information while enriching image views, ultimately enhancing the diagnosis of skin diseases. To validate the effectiveness of MCC, experiments were conducted by using ResNet-50, DensNet-169, Inception-v4, and ConvNeXt-B on both a self-collected hyperpigmented skin disease dataset and a public ISIC2018 dataset. The experimental results show that MCC can effectively improve the accuracy, precision, recall, and F1-score of CNNs. Thus, MCC has the potential to assist medical professionals in more accurately diagnosing skin diseases in clinical practice, thereby improving healthcare services and patients’ quality of life.