Multi-proportion channel ensemble model for retinal vessel segmentation

Objective: A novel supervised method that is based on the Multi-Proportion Channel Ensemble Model (MPC-EM) is proposed to obtain more vessel details with reduced computational complexity. Methods: Existing Retinal Vessel Segmentation (RVS) algorithms only work using the single G channel (Green Channel) of fundus images because that channel normally contains the most details with the least noise, while the red and blue channels are usually saturated and noisy. However, we find that the images that are composed of the alpha G-channel and (1-alpha) R-channel (Red Channel) with different values of a produce multiple particular global features. This enables the model to detect more local vessel details in fundus images. Therefore, we provide a detailed description and evaluation of the segmentation approach based on the MPC-EM for the RVS. The segmentation approach consists of five identical submodels. Each submodel can capture various vessel details by being trained using different composition images. These probabilistic maps that are produced by five submodels are averaged to achieve the final refined segmentation results. Results: The proposed approach is evaluated using 4 well-established datasets, i.e., DRIVE, STARE, HRF and CHASE_DB1, with accuracies of 95.74%, 96.95%, 96.31%, and 96.54%, respectively. Additionally, quantitative comparisons with other existing methods and cross-training results are included. Conclusion: The segmentation results showed that the proposed algorithm based on the MPC-EM with simple submodels can achieve state-of-the-art accuracy with reduced computational complexity. Significance: Compared with other existing methods that are trained using only the G channel and raw images, the proposed approach based on the MPC-EM, submodels of which are trained using different proportional compositions of R and G channels, obtains better segmentation accuracy and robustness. Additionally, the experimental results show that the R channel of fundus images can also produce performance gains for RVS.