retinalysis-fundusprep: A python package for robust color fundus image bounds extraction

TL;DR

This paper introduces an open-source Python package for automatic detection of bounds and contrast enhancement in color fundus images, improving accuracy and robustness for medical image analysis.

Contribution

The paper presents a novel, robust algorithm for CFI bounds detection and contrast enhancement, implemented as an open-source Python package, outperforming previous methods.

Findings

Error rate of 0.2% on EyeQ dataset

No mistakes on challenging Rotterdam images

Improved reliability over existing code

Abstract

Color fundus image (CFI) bounds detection and contrast enhancement are fundamental tasks in automatic CFI analysis. We present an open-source algorithm, published as a Python package, to automatically extract parametric bounds from color fundus images, and for applying contrast enhancement. The software has applications in automated biomarker calculation and AI systems. Bounds detection was implemented by detecting the CFI's contour via a shortest path algorithm on a polar transformation of the image. A second step detects points along the CFI circle robustly via a circle-fitting RANSAC algorithm. Straight boundaries are detected independently. Finally, the CFI is mirrored along its bounds before contrast enhancement to eliminate edge artifacts. We manually evaluated correctness on the EyeQ and Rotterdam Study datasets. Evaluation on the EyeQ CFI quality dataset revealed an error rate of 0.2% for our method, compared to 1.9% for previous work. On a sample of challenging CFI images from the Rotterdam Study, our method did not produce any mistakes. retinalysis-fundusprep is an improvement over existing publicly available code for CFI bounds detection. The improved reliability may allow for its use on a wider range of datasets without the need for manual validation.