Influence of background preprocessing on the performance of deep learning retinal vessel detection

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal vessel segmentation is crucial for diagnosing eye conditions and tracking retinal changes.
• Manual segmentation is labor-intensive and error-prone, necessitating automated solutions.
• Uneven image backgrounds in fundus photography pose a significant challenge for automated segmentation.

Purpose of the Study:

• To develop and evaluate an automated deep learning framework for retinal vessel segmentation.
• To assess the impact of various pre-processing techniques on segmentation performance.
• To compare the effectiveness of N4 bias field correction and CLAHE on fundus images with uneven backgrounds.

Main Methods:

• A patch-based deep learning framework utilizing a modified U-Net architecture was employed.
• Four pre-processing conditions were evaluated: no processing, N4 bias field correction, CLAHE, and a combination of N4 and CLAHE.
• Bayesian statistical testing was used to compare the performance of different pre-processing methods across three public datasets.

Main Results:

• The deep learning framework achieved competitive segmentation results on benchmark datasets.
• CLAHE pre-processing demonstrated the most significant positive impact on segmentation performance.
• While N4 correction offered marginal benefits for extremely uneven backgrounds, CLAHE alone provided substantial improvements.

Conclusions:

• Deep learning provides an effective approach for automated retinal vessel segmentation.
• CLAHE is a highly beneficial pre-processing technique for improving segmentation accuracy in retinal fundus images.
• Pre-processing strategies should be carefully considered, with CLAHE being a primary choice for enhancing segmentation performance.