Enhanced Feature Detection in Non-Proliferative Diabetic Retinopathy via RGB Channel Separation

Image pre-processing is crucial in biomedical stage detection systems since it enhances significant features, supporting subsequent image segmentation and classification phases. Particularly, in Non-Proliferative Diabetic Retinopathy (NPDR), stage-defining characteristics are extremely small and image pre-processing is vital to enhance the features for accurate stage diagnosis. This paper discusses an RGB channel separation analysis carried out to identify how each channel caters to emphasizing NPDR features using a set of color fundus images. Several methods were tested manipulating channel intensities, and the best performing method was chosen via the results generated from a U-Net segmentation model which was trained for each tested method. The results were compared using the U-Net dice scores, IOU values, and the identified and unidentified pixel percentages of the difference image which was obtained by overlapping the U-Net output image and the ground truth image. The best performing method presented a dice score of 0.80 and an IOU value of 0.66 on average.