Nuclear Segmentation And Classification Of Colorectal Cancer Pathological Images Based On Multi-Task Learning

At present,colorectal cancer has accounted for 12.2%of newly diagnosed cancer cases in China,becoming the second largest malignant tumor after lung cancer.Pathological diagnosis is the gold standard for tumor diagnosis.By analyzing the shape,type,and number of nuclei in pathological sections,we can help doctors to determine whether the tumor tissue of a patient has cancer cells and the malignant degree of the tumor,and guide clinical diagnosis.To analyze the nuclei in the pathological sections,accurate segmentation and classification is the most critical step.Although there have been many studies and reports on the nuclear segmentation and classification of colorectal cancer pathological images,the prediction performance of nuclear classification based on single cell segmentation heavily depends on the accuracy of the previous segmentation results.At present,most segmentation methods based on morphological methods have poor segmentation results for overlapped and adhered cells.In addition,there is currently a lack of high-quality datasets for pathological images of colorectal cancer.In view of these problems,this paper has carried on the following research work.First,to address the lack of a colorectal cancer pathology dataset,we manually labeled a new colorectal cancer digital pathology FFPE dataset,Seg CL.This data set included 51 digital pathologic FFPE sections of colorectal cancer from 15 patients,in which all nuclear information relevant to pathological prognostic features was accurately labeled by an experienced pathologist.The dataset contained a total of 29,002 nuclei,including 10,629 tumor cells,4,473 stromal cells,6,414 immune cells,4,524 necrotic cells and 2,941 other cells.Secondly,for colorectal cancer segmentation and classification problem,this paper proposes a new multi-task learning model De Sup-Net based on multi-task learning network Hover-Net by improving its encoder structure.De Sup-Net introduces dense residual connection in the encoder,extracts the deep feature information through local feature fusion and local residual learning,and integrates the tasks of nuclear segmentation and classification in digital pathological images of colorectal cancer into a unified network for synchronization,thus realizing the classification of nuclear pathological features while segmenting and identifying the nuclei.Finally,in order to further utilize the advantages of different network models,based on the integrated learning method,this paper integrates the results of Hover-Net,Micro-Net and De Sup-Net three models,and further improves the performance of the nuclear segmentation and classification task.Besides,in the prediction result,five types of nuclei such as tumor cells,immune cells and matrix cells are delineated by using different colors,so that the visualization of the prediction result is realized,and the prediction result is beneficial to efficient evaluation by pathologists.The experimental results show that the proposed De Sup-Net network has a DICE value of 0.885 and a PQ value of 0.711 for colorectal cancer cell nucleus segmentation.Compared with multiple existing cell nucleus segmentation methods,the proposed method has the best performance.The average F₁value for the five nuclei including tumor cells and immune cells is 0.838,with high classification accuracy.In addition,the classification accuracy of the ensemble learning model for tumor cells and immune cells is more than 0.5 points,and the accuracy is further improved,which is of great clinical significance for the grading judgment of colorectal cancer by pathologists.