Research On OCT Image Choroidal Segmentation Algorithm Based On Deep Learning

Choroidal segmentation represents a crucial component in medical analysis,particularly in the diagnosis and treatment of ophthalmic diseases.Although optical coherence tomography(OCT)has the ability to capture cross-sectional retinal images,it is still a challenging task to develop reliable and effective retinal segmentation methods owing to its shallow penetration depth,low image contrast,and heterogeneous choroidal texture.Furthermore,segmentation algorithms entail high data annotation costs,and currently,the research on deep learning approaches in this field is more oriented towards utilizing labeled data while little attention is given to unlabeled data usage.This factor may restrict the progress of algorithm accuracy and robustness.This article focuses on the utilization of semi-supervised learning methods to segment OCT images of the choroid,taking into consideration prior knowledge of the choroid structure.The paper details a series of studies conducted,outlining the specific tasks and research results achieved.(1)A cascaded segmentation network,known as the PGKD-Net,has been proposed to address the challenges associated with low contrast and the difficulty in accurately locating the boundary of the choroid membrane in OCT images.This network integrates prior knowledge of the choroid membrane to enhance segmentation performance.The segmentation process of the choroid membrane structure has been carefully designed.Experimental results demonstrate that utilizing a priori mask leads to a 1.79% increase in the Dice coefficient compared to not using such a mask.Furthermore,when the supervision weight of the priori mask in the loss function is set to 0.5,the model achieves its optimal performance,attaining a Dice coefficient of92.16%.This performance surpasses that of the current mainstream segmentation networks.Additionally,a novel module called the Multi Scale Context Aggregation Module(MSCA)has been proposed to enhance the model’s receptive field and context reasoning capabilities by employing deformable transformer capture.Experimental results indicate that incorporating the fusion of multi-scale context aggregation modules into the OCT dataset results in a 2.13% improvement in the Dice coefficient compared to the baseline network.(2)Considering the high cost and time-consuming nature of data annotation in practical applications,a semi-supervised learning framework is proposed.This particular framework leveraging Fourier transform to enhance image data and reduce distribution differences between various source images.Additionally,a confidence module is designed to assess the confidence of the segmentation results generated by unlabeled data.The evaluation results are incorporated into the loss of consistency regularization to minimize the impact of prediction errors on the model gradient.Experimental results demonstrate that this semi-supervised approach achieves comparable performance to fully supervised UNet and PGKD-Net at 20%and 50%labeling ratios,respectively.It is also shown that this framework yields a Dice coefficient of 92.78%on the 100%dataset.The approach described in this study has demonstrated promising outcomes in choroidal segmentation tasks and offers a potential solution to the issue of data annotation.These findings hold significant promise for enhancing the effective use of medical resources,enhancing diagnostic precision,and augmenting the diagnostic capabilities of physicians.