SD-OCT Retinal Geographic Atrophy Lesion Segmentation Based On Deep Weakly Supervised Learning

For the clinical diagnosis method,complete segmentation of retinal lesions is a critical clinical tool.Geographic atrophy(GA)is an advanced stage of non-exudative Age-related Macular Degeneration(AMD),which is an important leading cause of irreversible central vision loss.Benefiting from the high axial resolution,SpectralDomain Optical Coherence Tomography(SD-OCT)can provide a 3D perspective of the retinal structure.Deep learning has achieved satisfying performance in SD-OCT image segmentation.However,the dependence on a large amount of labeled data is still a major challenge in the application of deep learning technology.In this paper,the main approaches are based on the deep weakly supervised learning.Two models are proposed to operate the complex task of GA lesion segmentation by using more efficiently available annotations:(1)The Multi-Scale Class Activation Map(MS-CAM)is proposed by introducing the multi-scale fusion strategy into the Class Activation Map(CAM).A Scaling and Up Sampling(SUS)module is designed to balance the information between multi-scale features.An Attentional Fully Connected(AFC)module is proposed by introducing the attention mechanism into fully connected operations to integrate more effective features,where more important features are enhanced and redundant features are weakened.Based on the proposed MS-CAM,the information of GA positioning can be obtained for GA segmentation.(2)The Mirrored X-Net(MX-Net)is proposed to segment GA by joint the classification task and contrastive learning task.Adapting to the dimensional difference of information in SD-OCT images,a novel downsampling method,named Anisotropic Downsampling(ADS),is proposed to generate feature maps with different attention to each dimension.In order to avoid the limitations of single-task learning in weakly supervised methods,a novel contrastive learning module is proposed to form a multitask learning system with a classification task.Based on the proposed contrastive learning module,an Anomalous Probability Map(APM)is generated to discriminate the distribution of normal and lesion tissues in the image,which can be refined to obtained the final GA segmentation masks.(3)The performance of the proposed models was evaluated on two datasets.The experimental results show that the methods in this paper can obtain satisfactory results only trained with image-level annotations,and even higher accuracy than existing fully supervised GA segmentation method.