Segmentation And Classification Of Colorectal Histopathologic Cells Based On Deep Learning Method

Colorectal cancer,also known as colorectal cancer,is a kind of gastrotestinal malignancy that is relatively common world wide with a rather high fatality rate.The onset of colorectal cancer is usually located in the epithelium of colorectal mucosa at the beginning.Because of its occult onset,patients often do not feel any obvious symptoms before its development to the advanced stage.Therefore,the therapeutic effect of colorectal cancer depends on early detection and early diagnosis,among which early resection of cancer is the only radical cure at present.If the disease can be found at an early stage,it will not only reduce the therapeutic difficulty,but also have a better prognosis.At present,biopsy pathology under colonoscope is the gold standard for diagnosis of colorectal cancer.In hospitals,experienced pathologists usually complete the final diagnosis.With the continuous maturity of endoscopic technology and imaging technology,the quality of pathological pictures has been significantly improved.However,it is still difficult to achieve zero missed diagnosis and errors in examination and final diagnosis.Therefore,it is urgent to develop a computer-aided system that can accurately and quantitatively analyze pathological sections.As digital pathology technologies develop and computer hardware technologies develop rapidly,deep learning has gradually shown its unique advantages in the field of medical image processing.Hover-Net is a very effective network model for nuclear segmentation and classification of pathological images,which greatly facilitates the downstream analysis of computational pathology and the auxiliary diagnosis of colorectal cancer.Based on Hover-Net deep learning model and Lizard dataset of colorectal pathological pictures,this paper improves the original network by introducing attention mechanism and Swin transformer,and achieves better segmentation and classification results.The main contents of this paper are as follows:(1)The newly published semi-automatic labeling dataset Lizard focusing on colorectal tissues is used to train Hover-Net model,so as to obtain a prediction model with higher accuracy for segmentation and classification of colorectal pathological images.(2)To solve the problem of insufficient extraction of global feature information by convolutional neural network,Swin transformer,which has achieved great success in the field of computer vision,is introduced between decoder and encoder,and some details are changed to improve the prediction ability of the network.(3)Based on original Hover-Net model,different attention modules are added,such as SE,ECA,etc.,comparing the improvement effects of various attention modules on the model,and then selecting the attention module with less time cost and obvious improvement of segmentation and classification accuracy as the object for further improvement and optimization.