CF-Loss: Clinically-relevant feature optimised loss function for retinal multi-class vessel segmentation and vascular...

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Ophthalmology

Background:

• Accurate characterization of vascular features is crucial for diagnosing ophthalmic and systemic diseases.
• Existing methods for vessel segmentation may not directly optimize clinically relevant vascular metrics.

Purpose of the Study:

• To develop and validate a novel loss function (CF-Loss) that explicitly incorporates clinically relevant vascular features into the segmentation process.
• To improve the accuracy of multi-class vessel segmentation and the estimation of vascular features.

Main Methods:

• An end-to-end loss function, CF-Loss, was designed to categorize pixels into artery, vein, uncertain, and background classes.
• CF-Loss was integrated with standard segmentation networks and evaluated on three public datasets.
• The performance of CF-Loss was compared against standard segmentation metrics and its impact on a clinical downstream task (ischemic stroke prediction) was assessed.

Main Results:

• CF-Loss significantly enhanced both multi-class vessel segmentation accuracy and vascular feature estimation.
• Pixel-based segmentation performance did not always correlate with the accuracy of vascular features, underscoring the benefit of direct feature optimization.
• Improved vascular features derived from CF-Loss demonstrated quantitative benefits in predicting ischemic stroke.

Conclusions:

• CF-Loss offers a superior approach for optimizing vascular feature extraction through direct integration into the segmentation loss function.
• This method holds promise for advancing biomarker discovery and improving diagnostic accuracy in various diseases.
• The developed technique provides a valuable tool for clinical applications, particularly in predicting neurological conditions like ischemic stroke.