Privileged Modality Guided Network for Retinal Vessel Segmentation in Ultra-Wide-Field Images

Retinal vessel segmentation in ophthalmic images is an essential task to support the computer-aided diagnosis of eye-related diseases. As a non-invasive imaging technique, ultra-wide-field (UWF) fundus imaging provides a large field-of-view (FOV) of 200. with full coverage of the retinal territory, making it a suitable modality for vessel analysis. However, imaging the large FOV may result in low-contrast vascular details and background artifacts, which pose challenges to the accurate segmentation of retinal microvasculature. To address these issues, a privileged modality guided multi-scale location-aware fusion network is proposed for vessel segmentation in UWF images. We first perform style transfer on the UWF images to generate the corresponding FFA image with higher contrast. Afterwards, we employ cross-modal coherence loss to segment the vessels guided by the FFA image. Additionally, a multi-scale location-aware fusion module is proposed and embedded into the segmentation network for reducing the boundary artifacts. Finally, experiments are performed on a dedicated UWF dataset, and the evaluation results demonstrate that our method achieves competitive vessel segmentation performance with a Dice score of around 78.13%. This indicates that our method is potentially valuable for subsequent vessel analysis to support disease diangosis.