การปรับปรุงคุณภาพของภาพถ่ายเรตินา

Abstract

According to the World Health Organization (WHO), eye diseases such as age- related macular degeneration, cataract, and glaucoma are the main cause of blindness in the elderly worldwide. Most currently diagnostic systems are based on a color retinal photography. However, the image may be unsatisfactory for reliable medical diagnosis due to their low quality such as low contrast or poor color which caused from camera properties, non-uniform illumination, and the experience of photographers. For this problem domain, we propose Color Balance and Contrast Enhancement (CBCE) based on the histogram specification method for improving the image quality to provide better visibility of the retinal anatomical structures. The desired histogram is designed to enhance the contrast and color balance while preserving the naturalness of color retinal images based on Hubbard specification. It does this by employing Generalized Extreme Value (GEV) functions as transfer functions to redistribute the intensity. The optimal GEV parameters are determined by our image quality index namely Achromatic Contrast Sensitivity Quality Metric (ACSQM). The ACSQM is designed to deal with human visual system based on psychometric constraints and a contrast sensitivity function. The performance of our method has been evaluated against data from the STARE and DIATETDBO image databases. The average green-to-red and blue-to-red color balance ratios (Mean ± SD) of the enhanced images from DIARETDB0 are 0.506±0.002 and 0.166 +0.002, and STARE's ratios are 0.503 0.003 and 0.168 ± 0.002, which is close to the color balance specification from the Hubbard model. The contrast of green channel from DIARETDBO and STARE has increased by 141.68% and 21.63%. Moreover, the quantitative measure for LOE (18.17 14.82) shows that our method performs better than other methods in naturalness preservation. Subjective assessment by ophthalmologists shows that the contrast and color balance of the enhanced images have increased by 64.38% and 62.40%. The results obtained show that our CBCE algorithm performs well in the color retinal image enhancement. The enhanced images have better image contrast and color balance based on Hubbard model while retaining a pleasing natural appearance. In term of diagnosis by ophthalmologists, the enhanced images obtained by our method could be used to assist ophthalmologists in early detection and medical diagnosis.