AutoMorph Optimization for Enhanced Retinal Vessel Analysis in Premature Infants

Abstract

Purpose : Retinopathy of Prematurity (ROP) critically impacts the retinal vasculature in premature infants, necessitating precise vessel segmentation for effective diagnosis and treatment. This study aims to enhance vessel segmentation in ROP by optimising AutoMorph, a leading image analysis tool, on the in-house dataset. The goal is to provide ophthalmologists with a more accurate, reliable tool for diagnosing ROP from digital retinal images, focusing on the unique vascular changes characteristic of this condition, thereby improving early detection and intervention strategies.

Methods : AutoMorph, an image analysis framework, was fine-tuned for this study. The training involved 63 colour fundus images that were collected and annotated from Moorfields Eye Hospital, with a separate set of 4 images for validation and 7 for testing. We use the latest pre-trained model provided by the authors. To evaluate the model’s performance, we adopt commonly used metrics in the evaluation of blood vessel segmentation solutions. These metrics include Accuracy, Sensitivity, Specificity, Precision, F1, MSE, IOU.

Results : Segmentation performance was assessed using the Dice Score, a common segmentation accuracy metric, achieving a mean of 0.635 and a standard deviation of 0.060. For Artery segmentation on the test set, key metrics included 0.987 accuracy, 0.28 sensitivity, 0.997 specificity, 0.550 precision, 0.360 F1 score, 0.013 MSE, and 0.221 IOU. Similarly, Vesin segmentation metrics showed 0.984 accuracy, 0.645 sensitivity, 0.993 specificity, 0.718 precision, 0.675 F1 score, 0.016 MSE, and 0.511 IOU.

Conclusions : The fine-tuned AutoMorph demonstrated results in the segmentation of retinal vessels in ROP with a mean Dice Score of 0.635 and a standard deviation of approximately 0.060. The study underscores the potential of fine-tuning advanced image analysis tools like AutoMorph and highlights the efficacy of customizing advanced image analysis tools for specific medical imaging applications.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Visualisation of the testing set. The blue colour represents the Artery and the red colour for the Vein. The DSC are mean segmentation results over artery and vein.

 

Fine-tuned AutoMorph verses the grader’s annotation on the key metrics.