Multi-teacher based knowledge distillation for retinal vessel segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate retinal vessel segmentation is vital for diagnosing ocular diseases.
• Current methods struggle with segmenting thin vessels, impacting diagnosis and treatment.
• This limitation necessitates improved segmentation techniques for better patient outcomes.

Purpose of the Study:

• To develop a novel Multi-Teacher Based Knowledge Distillation (MTKD) method for Retinal Vessel Segmentation (RVS).
• To address the challenge of accurately segmenting both thin and thick retinal vessels.
• To enhance the robustness and accuracy of RVS compared to existing methods.

Main Methods:

• Proposed a Multi-Teacher Based Knowledge Distillation (MTKD) approach for RVS.
• Trained three specialized teacher networks focusing on original, thin, and thick vessel ground truths.
• Trained a student network to learn from the soft predictions of multiple teachers, minimizing knowledge discrepancy.
• Introduced a penalization technique to the student model's loss function for performance enhancement.

Main Results:

• The MTKD method demonstrated highly competitive performance on retinal fundus and angiography datasets.
• Improved a baseline U-Net model by up to 8.44 points in F1 score and 10.42 points in IOU.
• Ablation studies confirmed the effectiveness of the multi-teacher strategy, loss functions, and model complexity.

Conclusions:

• MTKD offers a promising approach for improving the accuracy and robustness of Retinal Vessel Segmentation.
• The method effectively learns to segment diverse vessel types by leveraging specialized teacher knowledge.
• Publicly available code and data promote reproducibility and further research in RVS.