Diagnostic Accuracy Of A Deep Learning Model In Grading Posterior Capsule Opacification
Published 2025 - 43rd Congress of the ESCRS
Reference: FP21.15 | Type: Free paper | DOI: 10.82333/wbr9-rq55
Authors: Eleonora Ferioli 1 , Linda Guo 1 , Alessandra Curci 1 , Giada Morese 1 , Giuseppe Giannaccare 2 , Filippo Lixi 2 , Claudia Corda 2 , Paolo Nucci 1 , Edoardo Villani* 1
1Università degli Studi di Milano,Milan,Italy, 2Università degli Studi di Cagliari,Cagliari,Italy
Purpose
To develop a deep learning model to automatically grade posterior capsule opacification (PCO) from retroillumination pictures.
Setting
This explorative study was conducted at the Department of Ophthalmology and Optometry at the Medical University of Vienna in cooperation with the Institute of Statistics and Mathematical Methods in Economics at the Vienna University of Technology (TU Wien).
Methods
A dataset of 355 retroillumination pictures from 271 pseudophakic eyes was subjectively graded by three experts regarding the severity of PCO on a scale of 0 to 10. Based on this gradings, five classes were defined: 0 (PCO grade 0), 1 (grades 1–2), 2 (grades 2.5–3), 3 (grades 3.5–4), and 4 (grades 4.5 and above). The dataset was split into 65% for training, 15% for validation, and 20% for testing to train a Convolutional Neural Network (CNN) for the grading task. The experiments were conducted using the original, unprocessed pictures. Model performances were assessed using sensitivity (recall) and precision.
Results
The CNN model developed for the PCO grading task showed sensitivity and precision as follows: 61.1% and 64.7% for class 0, 73.9% and 60.7% for class 1, 41.7% and 50% for class 2, 25% and 66.6% for class 3, and 70% and 53.7% for class 4.
Conclusions
The developed CNN model shows promising potential for objectively grading PCO, with further improvements expected by an image preprocessing process. Its potential as a reliable tool for both clinical and research purposes warrants ongoing development.