A Skin Cancer Detection Framework Based on Double-Branch Attention Neural Networks
Skin cancer is a major cancer and has increased rapidly in the past decades.Early detection can significantly increase the cure rate.Recently,deep learning models,especially various convolutional Neural Networks using dermatoscope images(i.e.,dermoscopy)have been widely adopted to classify skin lesions.Different from traditional image classification,several challenges in detecting and classifying skin cancers still exist,including imbalanced training data in each skin cancer category,small visual differences between categories,and small area of skin lesion.To solve these challenges,this paper proposed a skin cancer classification framework based on double-branch attention convolutional neural networks(DACNN).First,in data pre-processing,the whole dataset was divided into finer-grained categories according to the natural sub-classes in each category to alleviate the imbalanced data.Next,from the viewpoint of neural network structure,attention residual learning(ARL)modules were used as basic blocks in upper-branch,which was able to effectively extract the features of potential sick area,then the lesion location network(LLN)was designed to localize,cut out and zoom-in the sick sub-area,followed by being sent to down-branch with the same neural structure as the upper-branch,for extracting the locally detailed features.Then,the inferred features from both branches were integrated for effective detection and classification.Moreover,to further alleviate the impact of imbalanced categorical data,weighted loss function was utilized in the model training.The proposed DACNN model was implemented in the real dataset consisting of 10015 dermatoscope images and compared with several typical deep learning based skin lesion detection methods.Experimental results showed that the performance metrics of sensitivity,accuracy and F1_score reached 0.922,0.942 and 0.933,respectively.Compared with recurrent attention convolutional neural network(RACNN)detection methods,these three metrics were improved by 3.48％,2.95％and 3.44％respectively.In summary,our work significantly improved the accuracy of dermoscopy based skin cancer detection through appropriate division of dermatoscope image classes,used the double-branch attention neural networks to firstly localize and enlarge the features of potential sick area,and then further extracted the locally detailed features,which solved the intrinsic issues of dermatoscope images,including imbalanced samples in each skin cancer category,vague visual differences between categories,and small area of skin lesion.
skin cancerdouble-branch neural networkattention mechanismunbalanced data
NETL
NSTL
万方数据
