کنفرانس مهندسی برق ایران

صفحه اصلی / سی و دومین کنفرانس بین المللی مهندسی برق

Automated Optic Disc Segmentation in Low-Quality Retinopathy of Prematurity Retinal Images

نویسندگان :

Abolfazl Karimiyan Abdar¹ Reza AghaeiZadeh Zoroofi² Naser Shoeibi³ Sare Safi⁴ Alireza Ramezani⁵ Homayoun Nikkhah⁶ Hamid Safi⁷ Mohammad Reza Ansari Astaneh⁸

1- دانشگاه تهران 2- دانشگاه تهران 3- دانشگاه علوم پزشکی مشهد 4- دانشگاه علوم پزشکی شهید بهشتی 5- دانشگاه علوم پزشکی شهید بهشتی 6- دانشگاه علوم پزشکی شهید بهشتی 7- دانشگاه علوم پزشکی شهید بهشتی 8- دانشگاه علوم پزشکی مشهد

کلمات کلیدی :

retinopathy of prematurity،optic disc،segmentation،semi-supervise data augmentation،VGG-Unet

چکیده :

Premature infants are at risk of experiencing visual impairment primarily due to retinopathy of prematurity (ROP). Precise segmentation of the optic disc holds significant impact in determining the zone of ROP. Due to the imaging problems in premature infants and intricate nature of retinal fundus images, characterized by non-uniform illumination, low contrast between the background and the target area, the segmentation of the optic disc for infants poses a significant challenge, and there is limited literature reporting on this aspect. In addition to these challenges, the situation becomes more difficult when there are no annotations available. This paper introduces a method to tackle this issue by suggesting a semi-supervised dataset augmentation approach based on human feedback, aiming to enhance the performance of segmentation in images related to retinopathy of prematurity. VGG-Unet was set as the base model and these steps are iteratively implemented until further improvement in the result is unattainable. In this paper, two datasets were employed: (1) the publicly available Drishti dataset containing 101 fundus images from mature humans with corresponding annotations, and (2) private TMB dataset comprising 1054 images without any annotations. The VGG16-Unet model, when trained, faced challenges in effectively segmenting a specific dataset characterized by a significant distribution shift from the training dataset. Consequently, a method is required for segmenting TMB dataset without relying on expert retina specialists or annotated images. Our proposed approach aims to enhance segmentation performance by training the model on a public dataset and then applying it to the specific dataset. The first results without proposed method show the Jaccard score of 0.47 and Dice coefficient of 0.55. In proposed method after 3 epochs, we reach to the Jaccard score of 0.75 and accuracy of 0.85.