Research On Segmentation Algorithm Of Lesion Region In Medical Image Based On Weak Supervision

In recent years,with the rapid development and maturation of medical imaging and computer technology,automatic segmentation of medical image lesions has become increasingly important in clinical diagnosis.However,the accuracy of the pixel-level labels used in supervised training of these automatic segmentation algorithms largely depends on the clinical experience of doctors and has high annotation costs.This leads to a need for more pixel-level labels for medical imaging that impedes the practical application of automatic medical image segmentation.At the same time,there are many image-level class labels for medical images in clinical practice,such as labeling a medical image as diseased or normal.However,image-level labels only provide information on the category of medical images and cannot provide specific location information for lesion areas.Therefore,this thesis first trains a discriminative model using the category labels of medical images and further obtains lesion area information of diseased images.The thesis proposes a discriminative weakly supervised medical image segmentation algorithm based on mask fusion.This algorithm obtains the activation maps of lesion areas through three different ways:discriminative class activation map module,contrastive learning module,and internal attention of Vision Transformer(ViT).On this basis,the activation maps of lesion areas generated by multiple branches are complementary using mask fusion,and the fused lesion area activation maps are further refined using patch-level pairwise affinity.Finally,the segmentation mask of lesion areas is obtained through random walk.However,the multi-stage semantic segmentation model based on the discriminative model leads to bias between the training targets of a single stage and the overall segmentation target.Therefore,this thesis further proposes a generative weakly supervised medical image segmentation algorithm based on image domain transformation and mask contrast.This algorithm uses a generative image domain transformation network to transform images in the diseased and normal image domains,decoupling the lesion areas and normal areas of the input image and generating the activation map of the lesion area.At the same time,the discriminative class activation map module is introduced into the generative model framework,and contrastive and mutual supervision losses are introduced to supervise the activation maps of lesion areas with different properties(generative and discriminative)during the training process.The final segmentation mask of the lesion area is obtained by further fusion and using a dense conditional random field(Dense CRF).To solve the problem of data class dependence,this thesis also proposes a weakly supervised medical image segmentation algorithm based on contrastive learning,which only uses diseased image data to segment lesion areas.The model uses self-supervised contrastive learning to minimize the similarity between the vector representations of lesion area and normal area at the feature level and image level dimensions,while maximizing the similarity between the vector representations of lesion areas and normal areas.This method obtains the activation map of the lesion area and uses a dense conditional random field to obtain the final segmentation result.Lots of experiments on the BraTS brain tumor dataset,ISIC skin melanoma dataset,and COVID-19 dataset demonstrate that the segmentation performance of the proposed models in all aspects is superior to other advanced weakly supervised segmentation methods currently available.