Quantification of Fundus Autofluorescence Features in a Molecularly Characterized Cohort of More Than 3500 Inherited Retinal Disease Patients from the United Kingdom

Abstract

Purpose: To quantify relevant fundus autofluorescence (FAF) image features cross-sectionally and longitudinally in a large cohort of inherited retinal diseases (IRDs) patients. Design: Retrospective study of imaging data (55-degree blue-FAF on Heidelberg Spectralis) from patients. Participants: Patients with a clinical and molecularly confirmed diagnosis of IRD who have undergone FAF 55-degree imaging at Moorfields Eye Hospital (MEH) and the Royal Liverpool Hospital (RLH) between 2004 and 2019. Methods: Five FAF features of interest were defined: vessels, optic disc, perimacular ring of increased signal (ring), relative hypo-autofluorescence (hypo-AF) and hyper-autofluorescence (hyper-AF). Features were manually annotated by six graders in a subset of patients based on a defined grading protocol to produce segmentation masks to train an AI model, AIRDetect, which was then applied to the entire imaging dataset. Main Outcome Measures: Quantitative FAF imaging features including area in mm2 and vessel metrics, were analysed cross-sectionally by gene and age, and longitudinally to determine rate of progression. AIRDetect feature segmentation and detection were validated with Dice score and precision/recall, respectively. Results: A total of 45,749 FAF images from 3,606 IRD patients from MEH covering 170 genes were automatically segmented using AIRDetect. Model-grader Dice scores for disc, hypo-AF, hyper-AF, ring and vessels were respectively 0.86, 0.72, 0.69, 0.68 and 0.65. The five genes with the largest hypo-AF areas were CHM, ABCC6, ABCA4, RDH12, and RPE65, with mean per-patient areas of 41.5, 30.0, 21.9, 21.4, and 15.1 mm2. The five genes with the largest hyper-AF areas were BEST1, CDH23, RDH12, MYO7A, and NR2E3, with mean areas of 0.49, 0.45, 0.44, 0.39, and 0.34 mm2 respectively. The five genes with largest ring areas were CDH23, NR2E3, CRX, EYS and MYO7A, with mean areas of 3.63, 3.32, 2.84, 2.39, and 2.16 mm2. Vessel density was found to be highest in EFEMP1, BEST1, TIMP3, RS1, and PRPH2 (10.6%, 10.3%, 9.8%, 9.7%, 8.9%) and was lower in Retinitis Pigmentosa (RP) and Leber Congenital Amaurosis genes. Longitudinal analysis of decreasing ring area in four RP genes (RPGR, USH2A, RHO, EYS) found EYS to be the fastest progressor at -0.18 mm2/year. Conclusions: We have conducted the first large-scale cross-sectional and longitudinal quantitative analysis of FAF features across a diverse range of IRDs using a novel AI approach.

When Two Eyes Don’t Suffice—Learning Difficult Hyperfluorescence Segmentations in Retinal Fundus Autofluorescence Images via Ensemble Learning

“…From our analysis, we see that HF and particularly RA are difficult labels to segment precisely. Our mean results of 0.69 DcS agreement on HF and 0.36 DcS on RA are partly in accordance to the numbers recently reported by [7], who on FAFs with inherited retinal diseases had an inter-grader agreement of 0.72 DcS for HF and 0.75 DcS for RA.…”

“…To the best of our knowledge, which is supported by recent surveys [36][37][38] and research [39], there currently exists no public FAF dataset with annotated HF. Hence, like others [7,15], we can show results solely on in-house data, all of which were acquired with institutional review board approval from the University Eye Clinic of Kiel, Germany.…”

“…However, they often need training data with a clearly defined ground truth by human annotators. Now, if this ground truth is subject to the above-mentioned large inter-and intragrader variability, as is often the case in medical image segmentation [6], and which was very recently shown on FAFs with inherited retinal diseases [7], this leads to the consequent dilemma that the CNN has to deal with these ambiguities.…”

“…Their approach constitutes one of our baselines. Woof et al [7] very recently used a single nn-U-Net [17] ensemble consisting of five U-Nets to segment, inter alia, HF, and RA in FAFs with inherited retinal diseases.…”

“…Available sample sizes (less than 200 training images) due to an absence of stacked scans are also lower compared to [30][31][32] whose training set sizes, e.g., for lesions range from 1k to 16k images. To the best of our knowledge, and compared to methods like [7], we are the first to explicitly approach the problem of HF and RA segmentation for ambiguous FAF annotations.…”