Multi-proportion channel ensemble model for retinal vessel segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal Vessel Segmentation (RVS) is crucial for diagnosing eye diseases.
• Existing RVS methods primarily use the Green channel of fundus images, often neglecting valuable information in other channels.
• The Red and Blue channels are typically discarded due to noise and saturation, limiting segmentation performance.

Purpose of the Study:

• To introduce a novel supervised method, the Multi-Proportion Channel Ensemble Model (MPC-EM), for enhanced Retinal Vessel Segmentation (RVS).
• To leverage combined information from Red (R) and Green (G) channels to improve the detection of fine vessel structures.
• To achieve high-accuracy RVS with reduced computational complexity.

Main Methods:

• The proposed method utilizes an ensemble of five identical submodels, each trained on different proportional compositions of the R and G channels (αG + (1-α)R).
• Each submodel learns to capture distinct vessel features from these combined channel images.
• Probabilistic maps from all submodels are averaged to generate the final refined segmentation.

Main Results:

• The MPC-EM achieved high accuracies across four standard datasets: DRIVE (95.74%), STARE (96.95%), HRF (96.31%), and CHASE_DB1 (96.54%).
• Quantitative comparisons demonstrated superior performance compared to existing methods.
• Cross-training results validated the robustness and effectiveness of the proposed approach.

Conclusions:

• The MPC-EM offers state-of-the-art accuracy for RVS with significantly reduced computational complexity.
• This method effectively utilizes information from both R and G channels, outperforming traditional single-channel approaches.
• The findings highlight the potential of incorporating the R channel for improved RVS accuracy and robustness.