Deep-learning based quantification of RPE65-mutation inherited retinal degeneration from wide-angle images – novel biomarkers for natural history and interventional studies.

Abstract

Purpose : Subretinal gene therapy with Voretigene Neparvovec (VN) for inherited retinal degenerations with biallelic mutations in RPE65 (RPE65-IRD) maybe associated with accelerated retinal degeneration. Quantification of panretinal pathologies is an unmet need. We propose a deep-learning based algorithm to quantify progressive retinal changes in treatment naïve and treated patients using wide-angle color fundus photgraphs.

Methods : We trained a Mask R-CNN for automatized multi-class segmentation of distinct regions of interest (Fig. 1) representing different degrees of degeneration. We exploited a residual network for the underlying feature extraction with manual ground truth annotations for manual annotation on 150 Clarus 133° fundus images (17 patients, 34 eyes) and evaluated the performance of the algorithm on 32 Clarus 133° fundus images (6 patients, 12 eyes). Due to the relatively small size of the training dataset, we employed several data augmentation techniques including rotation, flipping, and adaption of contrast, intensity as well as color.

Results : Fig 2 shows pairs of ground truth annotations and segmentation predictions for bone spicules, spotty hyperpigmentations (top row), and marble-like pathology (bottom row). The proposed algoritihm exhibits a precision of 65.62% and a mean average precision 50 value of 67.3%, where insufficient segmentation masks are primarily due to the relatively small bone spicules and spotty hyperpigmentations.

Conclusions : The algorithm allows monitoring panretinal disease expression and progression based on the number and size of the predefined regions of interest and their spatial relation. Despite the use of a relatively small data set precision was acceptable. Transformer-type networks or the inclusion of synthesized training data using generative models can improve precision of the algorithm as well as including image sets from more patients. A larger cohort of RPE65-IRD patients was identified in the 2021 survey by the European Vision Institute for Research Net (EVICR.net) on the management of RPE65-IRDs in Europe.

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

 

Fig 1. Definition of six distinct fundus features in RPE65-IRDs (Clarus 133°, Zeiss Oberkochen, Germany)

 

Fig 2. Annotation of fundus features BS, SH and ML, and deep-learning segmentation with the novel algorithm.