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

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Axial misalignment in fundus imaging distorts anatomy, reducing diagnostic reliability.
• Disc-Fovea Line (DFL) orientation is crucial for rotation quality control (QC), but optic disc (OD) and macula localization is challenging.
• Limited pixel-level annotations hinder macula segmentation and data availability.

Purpose of the Study:

• To introduce FMSSL-GAN, an annotation-efficient Semi-Supervised Semantic Segmentation (SSSS) framework for fundus images.
• To improve the accuracy and efficiency of macula and optic disc segmentation for DFL-based QC.
• To provide an objective fundus image quality assessment solution.

Main Methods:

• Developed FMSSL-GAN, incorporating a Conditional Flow Matching (CFM) module for anatomical structure refinement via vector fields.
• Integrated a Generative Adversarial Network (GAN) to enforce topological consistency and spatial relationships between landmarks.
• Implemented an automated QC system flagging images with OD-macula angles exceeding 15°.

Main Results:

• The CFM-based approach enables faster inference with high-precision structural refinement compared to traditional diffusion models.
• FMSSL-GAN demonstrates robust performance across various labeled data ratios on private (Szeye) and public (REFUGE, ORIGA) datasets.
• The framework outperforms existing semi-supervised methods in fundus image segmentation tasks.

Conclusions:

• FMSSL-GAN offers an annotation-efficient solution for objective fundus image quality assessment.
• The method achieves high concordance with expert measurements for DFL-based QC.
• This framework enhances the reliability of clinical diagnoses by addressing axial misalignment issues.