Author: Dun Jack Fu (United Kingdom)
Co-authors: Gongyu Zhang, Bart Liefers, Livia Faes, Siegfried Wagner, Pearse Keane, Konstantinos Balaskas
Purpose
Geographic atrophy (GA) is the defining atrophic lesion of advanced non-neovascular age-related macular degeneration (AMD). Detection and segmentation of GA from optical coherence tomography (OCT) imaging is necessary for diagnosis, monitoring, prognosis, and to inform therapy research for this orphan disease. The current standard of segmentation requires specialist manual effort, which is labour intensive and prone to inter-grader variability. There is a need for validated and fully automated deep-learning approaches to qOCT detection and segmentation of GA that are applicable in clinical care of non-neovascular AMD patients with GA and potential to facilitate therapy research.
Setting/Venue
Dataset for model development were selected from participants of the FILLY2 trial (NCT02503332) a 18-month prospective, multicenter, randomised, sham-controlled phase 2 study that enrolled patients at 46 sites in the United States (New England Institutional Review Board, University of Miami, Mayo Clinic, Institutional Review Board of the Cleveland Clinic Foundation, Duke University Health System Institutional Review Board, and Research Compliance Office, Stanford University), Australia (Bellberry Ltd), and New Zealand (Northern A Health and Disability Ethics Committee, Health and Disability Ethics Committees, and Ministry of Health). External validation cohort were recruited from Moorfields Eye Hospital NHS Foundation Trust (London, United Kingdom).
Methods
Segmentation models were trained on FILLY2 participant data 5049 OCT B-scans from 399 eyes (200 patients) with GA secondary to AMD were manually segmented. Using U-Net architecture, deep-learning models were trained for each of the constituent morphological features of GA: RPE (retinal pigment epithelium)-loss; photoreceptor degeneration; and hypertransmission. A fourth model was trained for direct segmentation of GA i.e. overlapping regions of RPE-loss, photoreceptor degeneration, and hypertransmission. Predictive performance was validated on an external dataset recruited from a cohort-independent in time and geography from the model development cohort-from a retrospective cohort of all patients who attended Moorfields Eye Hospital NHS Foundation Trust as part of routine clinical care. The dataset comprised 884 OCT B-scans from 192 eyes (110 patients), of which: 151 eyes with GA secondary to non-neovascular AMD; 31 eyes with neovascular AMD ; 8 eyes with pigment epithelial detachment; 2 eyes with epiretinal membrane. Each OCT scan was annotated for the target features by two senior graders independently with adjudication by a senior ophthalmologist. The primary outcome is the agreement (DSC [dice similarity coefficient] and ICC [intraclass correlation coefficient) between model GA prediction and consensus of the 2 independent graders on the external validation dataset.
Results
When evaluated on the external validation dataset, the resultant models accurately segmented each of the three constituent features of GA: RPE-loss (median DSC [Dice similarity coefficient] ± SD [standard deviation] 0.95 ± 0.21), photoreceptor degeneration (0.96 ± 0.21), hypertransmission (0.97 ± 0.15), in addition to GA itself (0.96 ± 0.15). Model performance was greater than agreement between specialist human graders (RPE-loss 0.93 ± 0.31, photoreceptor degeneration 0.89 ± 0.20; hypertransmission 0.81 ± 0.30;GA 0.80 ± 0.30). Although the models were trained on annotated OCTs from patients with GA secondary to AMD, high performance was observed for each feature when considering OCTs also featuring other retinal pathologies including nAMD, PED, and ERM.
Conlusions
We report the development and validation of a deep-learning tool that can automatically process OCT scans for detection and quantification of GA that can be scaled across OCT volumes. The approach described: (i) performs fully automated segmentation of whole-volume OCT B-scans for GA; (ii) identifies and segments each of the retinal OCT features required to identify GA and its subtypes as per the CAM (Consensus of Atrophy Meetings) group consensus definitions; (iii) is trained and validated on the largest cohort size to-date; and critically, (iv) is externally validated on an independent dataset from real-world clinical practice. Here it is demonstrated that predictive performance equivalent to human specialist graders is robustly retained beyond the sample used for model development. Together, these features award this strategy the sophistication, resilience, and generalisability required to facilitate real-life clinical management of patients with GA, as well as, promote standardisation of clinical trial endpoints for GA research.
Financial Disclosure
Apellis Pharmaceuticals (Waltham, Massachusetts, United States) provided funding, input on the study design, and feedback on the manuscript.
Comments
-