Stage Detection in Retinopathy of Prematurity using Deep Learning for Ridge Segmentation
Sayed Mehran Sharafi1 , Shayan Pourmirbabaei2 , Afsar Dastjani Farahani2 , Nazanin Ebrahimiadib3 , Ramak Roohipourmoallai4 , Marjan Imani Fooladi5 , Elias Khalili Pour2 *
- Translational Ophthalmology Research Center, Farabi Eye Hospital, Tehran University of Medical Sciences
- Retinopathy of Prematurity Department, Retina Ward, Farabi Eye Hospital, Tehran University of Medical Sciences,
- University of Florida
- University of South Florida
- Clinical Pediatric Ophthalmology Department, UPMC, Children’s Hospital of Pittsburgh
Abstract: Retinopathy of Prematurity (ROP) is a retina disorder that mainly impacts preterm infants with lower weights and potentially leading to scarring and retina detachment. The diagnosis of ROP consists of 5 stages where early stages are critical for determining treatment choices. Stages 1-3 and normal retinas are more subtly classified by the existence, size, and shape of the demarcation line (or ridge, in later stages). Although human experts can readily recognize advanced stages of ROP, diagnosing the earlier stages (i.e. stages 1-3) is significantly influenced by inter-experts variabilities. Low-quality images characterized by focusing issues, contrast deficiencies, and uneven illumination negatively influence the accuracy of the diagnosis based on ROP images. In recent years, considerable efforts have been conducted to automate diagnosis of ROP through artificial intelligence to make the process more objective and accurate. In this study, we focus on segmentation of demarcation line, which is crucial landmarks in ROP diagnosis, through deep learning.
Methods: An encoder/decoder convolutional neural network (CNN) based on a modified UNet structure was utilized for the purpose of demarcation line segmentation. We used a pre-processing algorithm to enhance the contrast of the images and correct for uneven illumination artifacts. This study included 425 ROP images of 95 infants with the average birth weight of 1305±427 gr and average gestational age of 29.3±3 weeks. Using a GUI, an expert labeled the areas of demarcation lines in each image by creating a binary mask covering the demarcation line. By augmenting the dataset using geometric transformations, and cropping and resizing the images to 512*512 pixels, we obtained a dataset of 1700 images. To train our network, we utilized two subsets of 1350 and 350 images and their associated binary masks as training and validation data respectively.
Results: The proposed model underwent testing on 40 images from a different set of images and achieved a segmentation accuracy of 0.85 in terms of F1 contour matching score (BF).
Conclusion: Our research shows that using deep learning for segmentation of demarcation lines, provide a prerequisite for the automated diagnosis of ROP severity and enables reliable detection of the disease in its initial phases.
