A method for recovering shadow retinal OCT images by the auxiliary gradient branch model

Abstract: Optical coherence tomography has become the gold standard for diagnosing ophthalmic diseases. It obtains the retinal structure of the eye through noninvasive measurements. However, the absorption of incident light by blood flow in blood vessels makes shadow areas appear in B-Scan images at the vessel locations, which seriously affects the accuracy of retinal layer segmentation and the observation of small lesions. In this research, a two-stage neural network is proposed, which is established by auxiliary training head and a global attention mechanism between shadowed and non-shadowed areas. In the task of segmenting shadow areas, this work achieves that the segmentation mIoU of retinal B-Scan images is 79.07%. Based on the segmentation mask, the shadow area is accurately recovered. The SSIM of the Sobel edge image is used to evaluate the effect of the recovered area. The proposed method achieves an SSIMarea of 98.2% between the shadow areas after and before recovery, and the average pixel brightness is improved by 36.7 Compared to all existing methods, this work introduces model constraints to ensure the structural invariance of the original image for the first time, resulting in an improvement of SIMMtotal from 47% to 99.9%. Finally, the proposed method is an improvement of clinical retinal layer segmentation accuracy by comparing the effect of RPE-Choroid layer segmentation of four types of retinal B-Scan images.