An Choroidal Segmentation Network Based on Attention Mechanisms and Self-Calibrated Convolutions

The choroid, a highly vascularized tissue layer situated between the retina and sclera, plays a critical role in supplying nutrients and oxygen to the outer retina. The advancement of optical coherence tomography (OCT) technology has enabled high-resolution, non-invasive imaging of the choroid. Accurate segmentation of the choroid in OCT images is a crucial step for essential clinical insights for the diagnosis and treatment of retinal diseases. Manual annotation by clinicians remains the clinical standard for choroid segmentation in optical coherence tomography (OCT) images; however, this method is intrinsically susceptible to subjective variability and limited by operational inefficiency. We developed ASC-UNet, a novel UNet-based architecture optimized through attention mechanisms and enhanced convolutional operators, to address the challenges of choroid segmentation. Experimental evaluations demonstrate that ASC-UNET achieves superior performance compared to both the baseline UNet and existing variants, with statistically significant improvements in Dice coefficient (0.93 vs. 0.89) and IOU (0.87 vs. 0.81). Crucially, these advancements are attained without substantially increasing the parameter count or computational complexity, thereby preserving computational efficiency. The proposed framework demonstrates enhanced accuracy and reliability in automated choroid segmentation, offering a practical solution for clinical deployment while maintaining compatibility with resource-constrained medical imaging systems.