Enhanced Feature Representation of Retinal Fundus Images using Multi-Channel Fusion

Diabetic retinopathy (DR) is a disease that poses a global threat to human vision, and thus necessitates early detection. The current clinical screening procedure relies on ophthalmologists and is time-consuming. This paper focuses on enhancing the quality of fundus images to facilitate an efficient detection of DR. This work establishes the information-rich nature of the green channel in colored fundus images via an eigenvalue-based approach, followed by enhancing the quality of fundus images. Utilizing preprocessing on the green channel and grayscale representation, along with direct green channel extraction, a three-channel image is created. The proposed pipeline improves contrast and reduces noise, yielding high Structural Similarity Index (SSIM) and Edge Preservation Index (EPI) scores. The robust multi-channel fusion demonstrates potential for improving image features and reducing noise in established DR segmentation datasets, achieving a peak SSIM of 0.8939 and EPI of 0.9757 across different image fusions. The proposed method achieved improved measures as compared to the state-of-the-art, thus showcasing the efficacy of the proposed method in enhancing features of DR fundus images.