Detection of vitreoretinal interface abnormalities from OCT images based on DS-YOLOv7 network

Advances in optical coherence tomography (OCT) technology allow a clear view of the vitreoretinal interface (VRI). The abnormality of the VRI is one of the common symptoms of high myopia, mainly including posterior vitreous detachment (PVD) and epiretinal membrane (ERM). They can cause severe damage to the structure and function of the retina, leading to permanent vision loss. Therefore, fully automated detection of abnormalities at the VRI is crucial for the management of high myopia. This paper presents a DS-YOLOv7 network aimed at accurately identifying abnormalities, including partial PVD, complete PVD, and ERM from retinal OCT images. Built upon the YOLOv7 network, the proposed model integrates the advanced dynamic snake convolution (DSConv) module to capture the curvilinear characteristics of lesions, and the mixture of attention and convolution (ACMix) module to improve the precision and robustness of feature extraction through effective fusion of self-attention mechanisms and convolution. Moreover, the introduction of the efficient complete intersection-over-union (ECIoU) loss function further enhances the coordinate regression capability of the model. Three-fold cross-validation on a dataset with 1973 OCT B-scans from 46 patients shows that the DS-YOLOv7 achieved superior performance in vitreoretinal interface abnormality detection, with mAP@0.5 of 0.714, mAP@0.75 of 0.438, and mAP@0.5:0.95 of 0.424. The proposed model can provide an accurate and efficient diagnostic tool for patients with high myopia.