Research On Automatic Segmentation Of Rectal Cancer Based On Semi-supervised Deep Learning

In the world,colorectal cancer ranks third in the incidence rate of malignant tumors,and its mortality rate ranks second.With the development of medical imaging technology,imaging examination plays a more and more important role in clinical auxiliary diagnosis and treatment.Accurate segmentation of rectal cancer in MRI is the important stage in the clinical diagnosis,treatment and research.In recent years,the performance of deep learning-based image segmentation algorithm is far exceeding the classical image segmentation algorithms.However,the high cost of manual segmentation data limits the application of deep learning image segmentation algorithm in medical image segmentation task.In this paper,we propose a multi-task attention model MTA U-Net,and study the automatic segmentation method of rectal cancer based on semi-supervised deep learning to realize the automatic segmentation of rectal cancer tumor in DWI image with low labeling cost.By introducing attention mechanism,the inaccurate segmentation problems such as mis-segmentation and miss-detection in abdominal medical images are better solved.The mean teacher framework is used for semisupervised model training.By using unlabeled data to participate in model training,the labeled data required by the deep learning model is reduced.Aiming at the possible inconsistency of pseudo annotation information between classification and segmentation in semi-supervised training scenario,we propose a classification and segmentation consistency loss function,which effectively improves the accuracy of pseudo annotation information.In this study,brain tumor segmentation dataset BraTS 2021 and DWI rectal cancer dataset are used to verify the effectiveness of the method.The DWI rectal cancer dataset containing 62 labeled data and 205 unlabeled data.Compared with the fully supervised scenario,our model has better performance under 9 evaluation indexes on rectal cancer dataset,the segmentation IoU is increased by 2.20%.The experimental results show that our semi-supervised multi-task segmentation model based on mean teacher framework can effectively learn task-related information from unlabeled data to improve the model performance.