| The first step in medical image analysis is to identify the tissue,Therefore,for many medical analysis algorithms,background removal is a prerequisite to obtain effective and accurate results.Although it is very easy for an operator to identify the organizational area of WSI,it can be challenging for a computer,mainly because of the color changes and artifacts in WSI,moreover,it is difficult to detect tissues such as alveolar tissue,adipose tissue,and poorly stained tissue.Cancer area segmentation is the basis for pathologists to calculate the positive rate of cancer and analyze the symptoms of cancer,so the result of cancer area segmentation will directly affect the pathologists’ results.As one of the most difficult problems in medical image processing,cancer region segmentation has been favored by many researchers.For the segmentation of tissue foreground and cancer region,it is common practice to train two models separately to accomplish the corresponding segmentation task.For accurate segmentation result,a large amount of labeled data is needed to support,but the medical images,as sensitive data,are expensive to acquire and process.Based on the existing data sets,in order to improve the accuracy of the model as much as possible,a medical image segmentation algorithm based on multitask learning is proposed in this paper,to improve work efficiency and model accuracy.In view of the above problems,this thesis has researched the following three aspects:1.A LinkNet-based segmentation method of tissue foreground is proposed to extract tissue foreground from pathological images.Based on LinkNet network,the model is trained through image rotation,color perturbation,color channel change and other data enhancement methods to simulate pathological images.Compared with the traditional segmentation method,the proposed method not only has strong generalization ability,but also can get very good results in the IHC pathological image with weak staining and little difference between foreground and background.The experimental results show that the proposed method is far superior to the traditional segmentation method and can segment the tissue foreground accurately.2.A LinkNet-based cancer region segmentation method is proposed to extract cancer region from pathological images.This method is the same as the tissue foreground segmentation method,and uses the same training data,except that mask is labeled as cancer region.The experimental results show that the method used in this part of the cancer region segmentation is not ideal,and further research is needed.3.A multi-task learning segmentation algorithm based on improved LinkNet is proposed for tissue foreground extraction and cancer region segmentation in pathological images.The above two tasks are both segmentation tasks,and there are some commonalities,but training two models separately is a waste of time and cannot make the two models share information in the process of training.Based on the theory of multi-Task learning,this paper trains two models at the same time by modifying the network structure of LinkNet,adding a decoder block to LinkNet and re-designing a new loss function,so as to realize the information shared in training process.The experimental results show that the model trained by the method of multi-task learning has no significant change in the accuracy of tissue foreground region segmentation,but the accuracy of cancer region segmentation is improved greatly,so it has greater application value. |
