Research On Assistant Diagnosis Algorithm For Capsule Endoscope Bleeding Lesions Based On Deep Learning

Nowadays,gastrointestinal diseases are more common due to irregular lifestyle,and gastrointestinal bleeding is the initial symptom of many gastrointestinal diseases,which can not be underestimated.As a new way of digestive tract examination,wireless capsule endoscopy,which is safe,healthy,painless and noninvasive,brings convenience to patients,but the massive pictures taken during the work not only increase the work intensity of doctors,but also lead to missed diagnosis and misdiagnosis due to visual fatigue and other factors.Therefore,based on the images of capsule endoscopy,this paper studies the auxiliary diagnosis algorithm of gastrointestinal bleeding lesions.In this paper,the deep learning technology is used to assist the diagnosis of bleeding lesions from two aspects:firstly,the images of bleeding lesions are accurately classified from a large number of images;secondly,the images of bleeding lesions are effectively segmented.The main research contents are as follows:Firstly,the advanced ResNeSt model in deep learning backbone networks is applied to the classification of bleeding lesions,and the label smoothing method is used to modify the loss function to improve the classification effect.The experimental results show that the miss detection rate of this method for bleeding images is only 1.33%,which is 7.337%lower than that of traditional methods,and the false detection rate for normal images is only 0.47%.The classification speed on a single GPU is 275 pieces/s,and the performance is improved significantly.Secondly,the classic Unet model in medical segmentation is used to segment the bleeding lesions,which verifies the feasibility of automatic segmentation.In view of the poor segmentation effect,this paper improves the segmentation effect from two aspects:model structure and segmentation head structure:Firstly,according to the analysis of the structural defects of Unet and the characteristics of the sample data,six improvements are made to the model structure:1)The best up sampling method suitable for the bleeding segmentation task is selected in the experiment;2)A variety of new fusion structures are designed for the output features of the model;3)The traditional CBR module(conv+BN+relu)in the Encoder is replaced by ResNeSt network to solve the problem of insufficient feature extraction;4)Transfer learning method is used to alleviate the problem of relatively complex segmentation task and relatively less medical data;5)Reconstruction of ResNeSt feature extraction network based on Deformable convolution to solve the problem of irregular shape of bleeding area;6)Reconstruction of features extracted from Encoder based on Dilated convolution pyramid to solve the problem of different scale of bleeding area in data set.Secondly,in order to reduce the influence of redundant information on the features and further improve the segmentation performance,two improvements are made to the structure of the segmentation head:1)Self attention mechanism is introduced to reconstruct the features with low rank,and the Expectation Maximization Attention(EMA)module is introduced to optimize the computational complexity;2)On the basis of the Cross Entropy loss function,combined with Dice loss function to alleviate the imbalance between positive and negative samples in the data set.Finally,after the improvement of the above two aspects,this paper proposes DAFFRS-Unet-EMA model.The experimental results show that the pixel accuracy of the model reaches 92.7%,and the IOU reaches 84.92%;compared with the original Unet model,the pixel accuracy increases by 13.97%,and the IOU increases by 13.96%;compared with the current comprehensive strength of DeepLabV3+,the pixel accuracy increases by 2.08%,and the IOU increases by 3.32%;the segmentation effect is improved significantly.