Research On Classification Method Of Breast Pathological Image Based On Multilayer Perceptron And Residual Learning

Breast cancer is a high-risk tumor disease worldwide,and it is also accompanied by a high mortality rate.Medical experts advocate reducing mortality through "early detection,early diagnosis,and early treatment".The traditional manual reading diagnosis method is not only time-consuming and laborious,but the process is cumbersome and easy to cause misdiagnosis.In order to improve the diagnosis efficiency and accuracy of breast cancer,it is particularly necessary to use computer vision and artificial intelligence to assist the diagnosis of breast pathology.The pathological diagnosis results directly determine the doctor's medical plan,so the breast cancer pathological auxiliary diagnosis system must have high reliability.Compared with traditional classification methods,the automatic classification of breast pathological images based on convolutional neural network has greatly improved the operation complexity,diagnosis speed and accuracy.However,the pathological image has high resolution and manual labeling is very difficult.The existing data set is small and the model generalization ability is weak.In addition,the breast tissue itself has intra-class variability,and the complexity,refinement,and ambiguity of the pathological image morphology and texture features,making the model less recognizable to local features.Therefore,in view of the high cost and difficulty of obtaining breast pathological images,the intra-class variability of breast tissue itself,and the complexity,refinement,and fuzziness of the morphology and texture features of breast pathological images,this paper draws on the multilayer perception The strong ability of machine convolution to distinguish local features,combined with the idea of the residual network residual unit to simplify the learning process,deeply studied the multi-layer perceptual residual learning network model.The innovation of this article includes the following two aspects:(1)In view of the fact that the medical image data set is small and cannot meet the needs of network training,and the traditional data enhancement algorithm has the problems of mixing noise and causing the loss of important diagnostic information,this paper combines the sensitivity of the simple linear iterative clustering algorithm to the edge contour and Based on the distribution of gray values of image blocks,an optimization method based on simple linear iterative clustering data enhancement is proposed.Experiments show that this method is very helpful to improve the generalization of the network;(2)In view of the characteristics of breast pathological morphology,complex texture,refinement,small differences between classes,and multiple mutations within the class,this paper uses multi-layer perceptron convolution to distinguish strong local features,which is convenient for the network to capture into the field of view The deep useful information,combined with the residual network residual unit to simplify the learning process and enhance the excellent performance of gradient propagation,proposed a multi-layer perceptual residual learning method to improve the discrimination of local features.Experiments show that the model can abstract complex features in breast pathological images,and the classification accuracy can reach 96.3%.The simple linear iterative clustering data enhancement method combined with gray value distribution proposed in this paper can ensure the effectiveness of data enhancement more than geometric transformation and sliding window.The multi-layer perceptual residual model is more suitable for medical image classification tasks with complex background and fine texture features than VGG16 and residual networks.In the future,we may consider applying the multi-layer perceptron residual classification model proposed in this paper to the clinical breast pathology automatic diagnosis system.