Adversarial Vessel-Unveiling Semi-Supervised Segmentation for Retinopathy of Prematurity Diagnosis

TL;DR

This work proposes a semi supervised segmentation framework designed to advance ROP studies without the need for extensive manual vessel annotation, and offers a scalable solution for leveraging unlabeled data in pediatric ophthalmology, opening new avenues for biomarker discovery and clinical research.

Abstract

Accurate segmentation of retinal images plays a crucial role in aiding ophthalmologists in diagnosing retinopathy of prematurity (ROP) and assessing its severity. However, due to their underdeveloped, thinner vessels, manual annotation in infant fundus images is very complex, and this presents challenges for fully supervised learning. To address the scarcity of annotations, we propose a semi supervised segmentation framework designed to advance ROP studies without the need for extensive manual vessel annotation. Unlike previous methods that rely solely on limited labeled data, our approach leverages teacher student learning by integrating two powerful components: an uncertainty weighted vessel unveiling module and domain adversarial learning. The vessel unveiling module helps the model effectively reveal obscured and hard to detect vessel structures, while adversarial training aligns feature representations across different domains, ensuring robust and generalizable vessel segmentations. We validate our approach on public datasets (CHASEDB, STARE) and an in-house ROP dataset, demonstrating its superior performance across multiple evaluation metrics. Additionally, we extend the model's utility to a downstream task of ROP multi-stage classification, where vessel masks extracted by our segmentation model improve diagnostic accuracy. The promising results in classification underscore the model's potential for clinical application, particularly in early-stage ROP diagnosis and intervention. Overall, our work offers a scalable solution for leveraging unlabeled data in pediatric ophthalmology, opening new avenues for biomarker discovery and clinical research.

Figures (5)

• Figure 1: The domain difference among publicly available datasets (CHASEDB b27 & STARE b28) and our ROP dataset. CHASEDB primarily consists of images from adolescents and healthy populations, while STARE comprises images from adults and emphasizes heterogeneous pathology cases. In contrast, our ROP dataset is curated for infants' retinopathy of prematurity detection.
• Figure 2: The framework of our proposed semi-supervised learning method for vessel segmentation. The student network learns from the supervised and consistency loss. The vessel-unveiling prediction extracted from the uncertainty-aware vessel-unveiling module reveals the hidden vessels to ensure boosted student network's performance on unlabeled inputs. The discriminator takes latent representation of both labeled and unlabeled domain to close the domain gap by aligning the features.
• Figure 3: Uncertainty-aware vessel-unveiling Module. The vessel-unveiling reveals hidden vessels due to ROP image domain internal challenges discussed in section \ref{['sec:introduction']}. In the heatmap scale, the red color represents a higher value.
• Figure 4: Visualization result of all baselines and our model on ROP test set with 18 CHASEDB labeled data and 90 ROP unlabeled images during training. The ROP vessels by their underdeveloped structure and background tissue colors, are harder to recognize than public datasets. Our uncertainty-aware vessel-unveiling module and latent feature alignments help to extend the vessel extraction through the end of the retina.
• Figure 5: Visualization result of different methods on CHASEDB test set with 5 labeled data during training. To illustrate the discrepancy between other baselines, we set bounding boxes in the color red to share false positive regions.