Conditional Flow Matching-based Semi-supervised Segmentation for Quality Assessment of Fundus Images

Axial misalignment in fundus imaging can severely distort anatomical geometry, frequently reducing the reliability of clinical diagnoses. The orientation of the Disc-Fovea Line (DFL) remains the best method for rotation quality control(QC), but it can be challenging to locate the Optic Disc (OD) and macula, which weakens its utility. The weak textures and few pixel-level labels bring great difficulties for macula annotation, resulting in a scarcity of available annotation data. To address this issue, we intro duceFMSSL-GAN, anannotation-efficientSemi-Supervised Semantic Segmentation (SSSS) framework. In particular, a Conditional Flow Matching (CFM) module is added to turn rough, noisy predictions into anatomically correct struc = tures by learning a continuous evolutionary vector field. Our CFM-based approach directly models the probability path, which makesinference much faster while still keeping high-precision structural refinement. This is different from traditional diffusion models, which have slow, multi-step iterative denoising. A Generative Adversarial Network (GAN) is added to impose topological constraints at the distribution level, which enables the spatial relationship between landmarks to remain consistent. For the assessment, our system automatically marks the images as non-compliant when the calculated OD-macula angle is more than 15°. Extensive experiments on a private dataset (i.e., the Szeye) and two public datasets (i.e., REFUGE and ORIGA) demonstrate that the proposed method performs well with a wide range of labeled data ratios and outperforms other compared semi-supervised methods. Our framework provides an annotation-efficient solution for objective fundus image quality assessment, demonstrating a high concordance with expert measurements.