Research On Attention Network Method For Dermoscopy Image Recognition

The skin is the largest organ of the human body.In recent years,the number of patients with skin cancer has increased year by year.If detected and treated in time,it can be completely cured.In the medical field,the diagnosis of skin cancer is mainly based on the dermatologist’s observation of the shape and color at present,which is time consuming and laborious.Because of the high similarity between skin diseases,even a trained dermatologist can easily lead to misjudgment.With the development of artificial intelligence and computer vision technology,computer-aided diagnosis will be promising.Automated recognition from dermoscopy images is challenging.First of all,the most deadly of skin cancer is melanoma.Melanoma is similar to non-melanoma in dermatoscope images.Secondly,because the contrast is low between the diseased area and normal skin,the boundary is not obvious,and it is easy to ignore skin lesions.Finally,the dermoscopy images always contain hair,bubbles and ruler,which might be blur or occlude the skin lesions.To solve above problems,this paper has proposed an optimized algorithm for skin cancer recognition based on the previous research works.The main contributions of this paper as follows:Firstly,segmentation plays a decisive role for melanoma recognition in the step-by-step network.In order to reduce the interference of segmentation information,this paper has proposed adaptive attention network(SANet),which enables high-level features to guide the underlying network to automatically focus on the lesion area and reduce the loss of the underlying information.The experimental results show that the AUC is 0.8584,the accuracy is 0.8700,and the AP reaches 0.6365.Distinctive features are loss caused by single-scale features in traditional networks.In order to solve this problem,this paper has presented multi-scale attention network(PANet),including adaptive pooling,and selective the weight of feature maps.Experimental results show that the AUC is 0.8662,the accuracy is 0.8667,and AP reaches 0.6675.Secondly,to deal with inter-class similarity and intra-class variation of skin lesions,this paper has proposed deep metric learning,which constrains the feature vectors in the distance space(CL-loss)or angular space(Angloss),in order that the similar features are close,and different classes are far.Experiments show that the algorithm has the AUC of 0.8712,the accuracy of 0.8683,and the AP of 0.7018,which is better than others without any additional training data and other ensemble architectures in skin cancer recognition.