Research On Breast Cancer Recognition Method Of Whole Slide Images Based On Convolutional Neural Network

In the process of cancer treatment,early accurate diagnosis,localization and identification of tumors are the key to the subsequent treatment of tumors.This article focuses on the accuracy and speed of breast cancer detection,systematically analyzes and studies the segmentation accuracy and efficiency of deep learning technology in breast cancer detection of whole slide images,and proposes multi-resolution combined whole-chip prediction to realize digital breast tissue The diagnosis of the full-slice lesion area provides computer-aided diagnosis for manual reading.The main research works of are as follows:1.For the problem of low segmentation accuracy of Unet,an AC-Unet segmentation model is proposed,and three improvements are made on the basis of the traditional segmentation model Unet: 1)Reduce the pooling layer and slow down the loss of small pixel targets;2)Introduce atrous convolution Product,increase the receptive field of the network,and improve the segmentation accuracy;3)Introduce the residual structure and feature fusion method in ResNet network to strengthen the semantic feature information.Compared with the Unet model and TernausNet model of other methods,the AC-Unet model has achieved better results,and the mIoU value has increased by 10.5% and 8.81%,respectively.2.For the problem of segmentation efficiency when the segmentation model is actually used,a DS-AC-Unet segmentation model with high efficiency is proposed.Based on the high-accuracy segmentation model AC-Unet,three improvements are made: 1)Introducing deep separable volume Product,optimize network parameters,and increase the speed of the model;2)add a BN layer after the hole convolution layer to speed up the training of the model;3)improve the data amplification algorithm and enhance the generalization ability of the model.The experimental results of the improved model are analyzed.Compared with the AC-Unet model,the prediction speed is increased by 7.6ms / patch,the model size is reduced by 30.5%,and the segmentation efficiency of the improved DS-AC-Unet model is improved.The stacking structure of convolutional layers was optimized again,and a TAC-Unet segmentation model with faster prediction speed was proposed,which was 12.3ms /patch faster than DS-AC-Unet prediction speed.3.According to the proposed segmentation model,the lesion area detection is performed on the breast tissue at the full slice level.By using a combination of multiple resolutions,fast and accurate lesion area detection and identification is achieved.A resolution-related inquiry experiment was performed,and the effects of various models were compared.Finally,a full-slice lesion area segmentation system based on DS-AC-Unet was obtained.The average full-slice prediction time was 45.6s and mIoU reached 93.12%.Provide possibilities for engineering implementation and applications.