Research And Implementation Of Breast Cancer Histopathological Image Classification Based On Deep Learning

As the most prevalent malignant tumor among women in worldwide,early detection and diagnosis of breast cancer is important to provide treatment in time and curb the progression of the disease.However,the analysis of histopathological images based on manual recognition consumes a lot of manpower.The bias of pathologists’ subjective consciousness and the complexity of images also can affect the diagnosis results,even lead to misdiagnosis or omission.So the use of computer-aided histopathological image diagnosis by physicians is essential.Most of the current computer-aided diagnostic methods for breast cancer histopathological image classification tasks have some problems such as low classification accuracy,lack of detailed classification,excessive amount of model parameters,and lack of clinically oriented applicable systems.Recently,some researchers have introduced the Self-Attention mechanism into the image classification tasks to solve the limitation of small receptive field in convolutional operation and obtained good results.Inspired by it,this thesis combines the Self-Attention mechanism with convolutional neural network.On the one hand,it solves the limitation of convolutional operation,and on the other hand,it compensate for the limitation of the large collective data volume requirement of Self-Attention networks.Meanwhile,a model compression algorithm is proposed to prune the trained model to reduce the computation time and the number of parameters.Based on the above research,a complete classification system is designed for clinical application.The main works are as follows:(1)We propose a breast cancer histopathological image classification network that combines convolutional neural network and dual-scale Self-Attention mechanisms.We add advanced channel Self-Attention methods and spatial Self-Attention methods to the residual network.It overcomes the disadvantage that the convolutional neural network can only process features in image’s neighbor regions due to its localized attribute.At the same time,it also overcomes the over-reliance on a large number of training images alone.In the training stage,an adaptive-TTA strategy is proposed for breast cancer histopathological images.We trained the network by using publicly available dataset Break His,and the effectiveness of the classification network is verified by experiments.The results show that the proposed classification model is able to get a high accuracy.(2)In order to improve the computational speed of the model in practical applications,a pruning method based on the SE mechanism and upper quartile truncation is proposed to compress the model.The SE mechanism can evaluate the importance of the model channels,we use it and select the upper quartile as the truncation factor to delete the unimportant filter channels and prune the trained optimal model.This process reduces the amount of parameters while ensuring high accuracy,and improves the generalization ability of the model for better application in computer-aided diagnosis in clinical application.(3)Based on the proposed breast cancer histopathological image classification network,we use the compressed model as the basis for classification.According to the functional requirement analysis,a breast surgery information management and image classification system is designed.It is mainly divided into four modules: new case,view historical cases,quick diagnosis,and system management.The system can store basic patient information and diagnosis records in the database for management and query.It achieves to show the real-time results and the probability of each disease output by the classification model to doctors.The compressed model is smaller in size and more reduction in computation,which makes the system work well on devices with low hardware configuration and has a strong clinical applicability.