A double-pass fundus reflection model for efficient single retinal image enhancement

Abstract

Purpose : High contrast is a prerequisite for fundus images to be used for successful diagnosis in clinical practice. Current contrast enhancement models show promising results, but they have not been optimized for retinal images, due to the ignorance of specific double-pass fundus reflection feature. In this study, we propose the double-pass fundus reflection (DPFR) model and test the hypothesis that the DPFR model can be more efficient and simpler for retinal image enhancement than previous methods.

Methods : The DPFR model is derived according to the light propagation in a fundus camera. Based on the model, the dark-channel prior de-hazing is applied twice to a raw retinal image to correct the uneven illumination, and reduce the straylight effect of the image. Then we convert the de-hazed image into CIE-Lab color space to apply the Contrast Limited Adaptive Histogram Equalization to its L channel, resulting in the final enhanced image. We compared the performance of the DPFR model with two state-of-art methods including the Luminosity and Contrast Adjustment (LCA) and Pixel Color Amplication (PCA). All methods are applied to the data from four public databases, namely the DRIVE-, the STARE-, the DiaRet- and the Cataract database. Both visual and objective assessments are shown. For objective assessments, we use the multiscale image contrast, C_RAMM, and the color difference between raw images to evaluate enhancement performance.

Results : Figure 1 shows that the DPFR model effectively balances the uneven illumination, and improved the clarity of retinal images since structures like blood vessels that were hidden behind the cataract in the raw image could be observed after enhancement. Table 1 shows the mean value of C_RAMM and color differences of raw and enhanced images from the four separate databases for the LCA, PCA, and DPFR methods, respectively. We found significant differences between the three methods (all P < 0.001). Although the PCA resulted in higher C_RAMM than other methods, it has worse color preservation (larger color difference value), as can be seen in Fig. 1, which makes this method less suitable for clinical use.

Conclusions : The DPFR model enables efficient illumination correction, contrast improvement, and color preservation. It on average enhances image contrast by two-fold, while the color differences are less than 10%, and show better performance compared to existing methods.

This is a 2021 ARVO Annual Meeting abstract.

 

Figure 1

 

Table 1