Retina-TransNet: A Gradient-Guided Few-Shot Retinal Vessel Segmentation Net.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• High labor costs for physicians limit manually-labeled medical image datasets for developing computer-aided diagnosis (CADx) and segmentation algorithms.
• Developing effective segmentation algorithms for medical imaging, such as retinal vascular segmentation, is crucial for disease diagnosis and monitoring.

Purpose of the Study:

• To develop a retinal vascular segmentation network with strong cross-domain generalizability using limited training data.
• To address the challenge of insufficient labeled medical images by reframing segmentation as an image-to-image (I2I) translation problem.

Main Methods:

• Proposed an image-to-image (I2I) translation framework for retinal vascular segmentation.
• Introduced a two-stage Unet (2Unet) generator with skip connections, where the second stage acts as a refinement module.
• Incorporated gradient-vector-flow (GVF) loss constraints to enhance the segmentation accuracy and generalizability.

Main Results:

• The proposed I2I translator-based segmentation subnetwork demonstrated superior cross-domain generalizability compared to existing methods.
• The model achieved stable segmentation results on diverse datasets (CHASE-DB1, STARE, HRF, DIARETDB1) when trained on a single dataset (DRIVE).
• Effective performance was maintained even in low-shot learning scenarios, indicating data efficiency.

Conclusions:

• Recasting retinal vessel segmentation as an image-to-image translation problem is an effective strategy for improving cross-domain generalizability.
• The proposed 2Unet generator with GVF loss enables robust segmentation with limited training data, offering a valuable tool for medical image analysis.
• This approach mitigates the need for extensive manual labeling, reducing physician labor costs in CADx system development.