Research On Weakly Supervised Biomarker Segmentation With Image-level Annotation In OCT Images

Segmentation of biomarkers from OCT images is a critical step in the diagnosis of retina-related diseases.Although fully supervised deep learning models can segment pathological regions,their performance relies on labor-intensive pixel-level labeling.Compared to intensive pixel-level annotation,image-level labeling can reduce the burden of manual annotation.However,existing weakly supervised segmentation methods based on image-level labeling still suffer from model collapse,training instability,and anatomical mismatch due to the large variation in shape,texture,and size of retinal biomarkers.In order to effectively reduce the annotation burden of ophthalmologists in OCT images,this thesis proposes two methods for the retinal biomarker segmentation problem based on image-level labeling,respectively.(1)A weakly supervised segmentation method for biomarkers based on anomaly localization is proposed.Specifically,in this thesis,the retinal biomarker region is considered as an abnormal region different from the normal area.First,to encourage the model to learn the anatomical structure of normal OCT images,a novel pre-training strategy based on supervised contrastive learning is proposed in this thesis.Second,to preserve the anatomical structure of the retina and enhance the coded representation of retinal features,a novel hybrid network structure is proposed in this thesis.The network includes supervised contrastive loss for feature learning and cross-entropy loss for classification learning.In addition,an efficient strategy is used to combine these two losses.Finally,to alleviate the challenge of insufficient supervision,an anomaly segmentation method based on knowledge distillation is designed in this thesis,which is effectively combined with the previous network to further improve the model performance.Experimental results on both local and public datasets demonstrate the effectiveness of the method.(2)A weakly supervised segmentation method for biomarkers based on Transformer and prototype learning is further proposed.Due to the local nature of the convolution operation,no remote dependencies can be established between the features extracted by the model.Considering that retinal biomarkers are closely related to their anatomical structures,the long-range information contained in OCT images may also be helpful for biomarker segmentation.The Transformer model for sequence prediction has a global self-attentive mechanism that can establish strong remote dependencies,but its inability to provide detailed low-level feature information may affect the localization ability of the model.In this thesis,we propose a novel network structure combining convolution and Transformer,which has the advantages of both convolution and Transformer and can effectively extract biomarker-related features from OCT images.To further alleviate the challenges caused by insufficient supervision and allow the model to focus more on biomarker regions,a prototype-based learning framework is proposed in this thesis to help the model capture biomarker features more effectively.In addition,to further improve the segmentation results,a multi-scale feature fusion module is proposed in this thesis to achieve more effective localization by combining different scales and types of feature maps.The proposed method is validated on a local dataset,and its experimental results show that the proposed method performs well on weakly supervised segmentation of biomarkers.