MTD-Net: A robust multi-task discriminative network for choroidal neovascularization segmentation

<p dir="ltr">Choroidal neovascularization (CNV) is a major manifestation leading to blindness in individuals aged 50 and above worldwide. The precise delineation of CNV in optical coherence tomography angiography (OCTA) images plays a critical role in the clinical diagnosis of age-related wet macular degeneration (AMD). However, due to the complexity of CNV shapes (irregularities) and limitations imposed by imaging factors such as projection artifacts and noise-induced boundary blurring, automatic segmentation of CNV faces significant challenges. Therefore, we propose a robust multi-task discriminative network to achieve precise segmentation of CNV lesion areas and vessels. Specifically, to address the intricate shapes and significant variations in CNV scales, we adopt a multi-task approach to learn different features of lesions, where a neural discriminative approach is designed to discerningly fuse information between different tasks. This involves integrating advantageous information from tasks like edge regression and shape learning to enhance the efficiency of region and lesion segmentation. Additionally, we introduce an uncertainty estimation strategy to reinforce the network’s ability to discern fuzzy boundary pixels, effectively addressing the issue of blurred edge pixels caused by artifacts and noise. Experiments on the collected CNV dataset show that our method outperforms existing medical image segmentation methods and CNV segmentation techniques. The Dice coefficients for region segmentation and vessel segmentation achieve 91 % and 90 %, respectively. In conclusion, this study introduces a robust segmentation framework for accurate CNV lesion and vessel localization, supporting improved diagnosis, treatment planning, and disease monitoring in clinical ophthalmology.</p>