Skin Disease Image Analysis Based On Multi-Task Learning

Skin disease is one of the diseases with the highest incidence in human beings,and has become a major public health problem affecting the economic and social development of the country.Early diagnosis and treatment of skin diseases is very important,which can greatly reduce the harm caused by the later stage of the disease,but it faces the problems of uneven distribution of medical resources and imbalance of doctor-patient ratio.Therefore,the use of artificial intelligence technology to achieve computer-aided diagnosis of skin diseases has great and far-reaching significance.Automatic classification and segmentation of skin disease images are two key tasks for realizing intelligent skin disease diagnosis.However,the existing methods are all designed based on dermoscopic images,the image data is not easy to obtain,and they are usually only applicable to one task,ignoring the correlation between the two tasks.In addition,there are certain performance problems,such as the lack of focus of classification features and blurred boundaries of segmentation results.In response to the above problems,this paper first proposes two clinical dermatology image classification and segmentation datasets for follow-up research.Secondly,this paper proposes a skin disease classification and segmentation network based on multi-task learning,aiming to mine the association between classification and segmentation tasks.The network introduces an auxiliary task of boundary prediction and consists of five parts: encoding network,segmentation decoding network,boundary decoding network,boundary aware segmentation module and classification sub-network.The encoding network provides common feature representations for the three subtasks.The segmentation and boundary decoding network consists of a set of attention decoding units based on an interactive attention mechanism,aiming to focus on the effective features in the decoding process.A set of boundary aware segmentation modules based on boundary attention are designed between the segmentation decoding network and the encoding network,aiming to introduce boundary feature information to enhance the boundary details of segmentation features.Two region of interest feature attention modules are introduced between the classification sub-network and the encoding network,aiming to use the segmented lesion region probability map to pay attention to the lesion region features in the classification features and weaken the influence of interference features.Finally,extensive experiments are conducted on the two datasets proposed in this paper.Compared with the most advanced classification and segmentation networks,the multi-task learning network in this paper achieves the best classification performance and segmentation performance.Through the ablation experiment of these modules,the effectiveness of each module in the multi-task learning network is verified.