Automated Segmentation of Geographic Atrophy Using Machine Learning on Real-World Fundus Autofluorescence Ophthalmic Images in the IRIS® Registry

Abstract

Purpose : To develop a machine learning (ML) pipeline to automatically segment and assess the size of atrophic lesions in patients with geographic atrophy (GA) using real-world fundus autofluorescence (FAF) images linked to the American Academy of Ophthalmology IRIS® Registry (Intelligent Research in Sight).

Methods : The IRIS Registry is the nation’s first comprehensive ophthalmic clinical database. A ML pipeline was developed using de-identified FAF images linked to the IRIS Registry. Images were linked to patients with GA, identified via ICD-10 codes and previously developed ML imaging models that identify patients with GA secondary to dry AMD from 2006 to 2022. Training, validation, and test sets were selected using patient-level stratified sampling (on image quality and lesion area, using a 60:20:20 split) and labeled by an expert grader. A deep learning model was trained on the training and validation sets to segment GA lesions. To assess the accuracy of the model on a patient-journey level, the model inference was applied to patients with multiple images over time, extracting the area of the lesion over time.

Results : A total of 214 FAF images from 214 unique patients were labeled. The training, validation, and test sets consisted of 122, 44, and 42 images respectively. A modified U-Net model was trained from scratch from the training set. The model reached a dice score coefficient of 0.861, 0.905, and 0.887 on the training, validation, and test sets respectively. To assess how the model could be used longitudinally, 20 patients not within the 214 labeled cohort with FAF images linked to the IRIS Registry were randomly selected. These patients had at least 2 imaging visits taken at least 3 months apart. From this patient cohort, all patients reported positive growth rates, with an average (SD) of 0.330 (0.34) mm/y, which were comparable to previously reported values.

Conclusions : The proposed pipeline demonstrates satisfactory accuracy for segmenting atrophic lesions in real-world FAF images. This developed ML pipeline could be leveraged to assess real-world GA disease progression at scale.

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