Research On OCT Image Biomarker Detection Based On Boundary Box Annotation

Retinal biomarker in optical coherence tomography(OCT)images plays a key guiding role in the follow-up diagnosis and clinical treatment of eye diseases.Although there have been many deep learning methods to automatically process retinal biomarker,the detection of retinal biomarkers is still a great challenge due to the similar characteristics to normal tissue,large changes in size and shape and fuzzy boundary of different types of biomarkers.To overcome these challenges,this thesis proposes two solutions for the detection of retinal biomarkers in OCT images by using bounding box annotation.(1)This thesis proposed a framework for detecting retinal biomarkers based on contrastive uncertainty,which only requires the annotation labeled by the bounding box to complete the training of the network,and uses the proposal contrastive learning to enhance the feature representation of retinal biomarkers.This strategy aims to improve the network’s ability to distinguish features between different types of retinal biomarkers.Thus,the network can more accurately classify different categories of retinal biomarkers.Furthermore,in order to enhance the sensitivity of the network to the fuzzy boundary of retinal biomarkers,a bounding box uncertainty strategy is added to the overall framework.Specifically,it combines boundary uncertainty with traditional bounding box regression to obtain better boundary location results.The performance of the framework was evaluated on a public dataset and a local dataset.The experimental results show that the framework has better performance in detecting retinal biomarkers than other object detection methods.(2)This thesis proposed a novel retinal biomarker detection method based on transformer in OCT images.The method is based on transformer encoder-decoder structure to detect retinal biomarkers in OCT images by better focusing on the association of retinal biomarkers with global information and using the retinal biomarker detection task as a direct ensemble prediction.In addition,due to the irregular shape of the retinal biomarkers,the deformable attention mechanism is introduced to enable the network to pay more attention to the location information,and make the network closer to the shape of the retinal biomarkers when selecting sampling points.Furthermore,due to retinal biomarkers contain some small area objects,in order to enable the network to better detect these objects,the multi-scale characteristics of swin transformer are utilized to enable the network to fuse information about small area retinal biomarkers in high-resolution feature maps,thus reducing the problem of information loss in low-resolution feature maps.The first method in this thesis mainly focuses on the classification and localization of retinal biomarkers,but the detection effect for smaller biomarkers is not ideal.The second method takes into account this issue and proposes a multi-scale strategy to solve it.Experiments have shown that the proposed methods effectively solve some of the problems encountered.