Research On Weakly Supervised Medical Image Segmentation Algorithm Based On Deep Generative Model

Medical imaging industry is the largest segment in the field of medical devices in China.In recent years,with the rapid development and substantial progress of artificial intelligence technology and medical imaging technology,it has become an essential technical means in clinical disease diagnosis,disease treatment and modern medical research.In the field of medical imaging-assisted diagnosis,biomedical imaging has become an integral and increasingly important part of disease diagnosis and treatment,as physicians or researchers often need to know detailed information about internal tissues and organs in order to be able to make the best possible treatment decisions when performing qualitative and quantitative analysis,real-time monitoring of conditions,and future treatment planning.Current auto-segmentation methods of medical images are greatly hampered by the insufficient and ambiguous clinical annotation.Actually,the rough classification labels（e.g.disease or normal）rather than the precise segmentation mask,are more common and available in clinical practice,but how to fully utilize those weak clinical labels to infer the precise lesion mask and guide the medical image segmentation remains largely unexplored.Compared with fully supervised semantic segmentation methods,the annotations required for weakly supervised semantic segmentation methods are easily accessible and less costly.Many excellent weakly supervised semantic segmentation works have emerged in recent years,and these methods have achieved good results in natural images,medical imaging,remote sensing images,and other fields.Weakly supervised signals include image category annotation,bounding box annotation,manual scribble annotation,and marker point annotation.Different weakly supervised semantic segmentation works have tried to investigate the trade-off between test accuracy and training data annotation from different perspectives.Among these types of supervised signals,the image categories have the weakest supervisory strength,and the image categories only tell what objects are in the image,but not which pixels belong to these categories.However,category annotation is the least costly and the most obtainable,so many works using only category information as a supervisory signal have emerged in the field of weakly supervised semantic segmentation,and we classify and organize each of these works in the main paper.In this paper,we proposed a weakly supervised medical image segmentation model to directly generate lesion mask by a class attention map（CAM）guided cycle-consistency label-activated region transferring network.Our model is based on a cycle generation adversarial network architecture,to which a CAM module is added to generate an initial lesion segmentation region,called the label activation region.To eliminate the blurry boundary of CAM-guided segmentation,we introduced a complementary branches fusion module to focus the transformation only on the label-activated region.The complementary branch fusion network contains three branches,namely the image transformation branch,the lesion mask branch,and the reconstruction branch.The feature maps of the three branches serve to complement each other and are fused in the corresponding channels.This model can simultaneously identify pixel level label-discriminated mask meanwhile maintain the semantic information and anatomical structure of medical image to precisely define segmentation boundary by collating the label and semantic related region through an image synthesis procedure.In addition,the produced lesion mask is further prompted by a joint discrimination strategy.Comparative experiments and ablation experiments on the BraTS dataset,the ISIC dataset,and the COVID-19 dataset show that the segmentation performance of our method outperforms other existing state-of-the-art weakly supervised segmentation methods in all aspects.In addition,we also conduct lesion segmentation experiments on the small sample dataset,demonstrating that our model can achieve excellent segmentation results using only a small amount of sample data for training.The analysis based on this series of experiments also demonstrates the effectiveness of our model.