Retinitis Pigmentosa Visual Impairment Predictable Using Deep Learning

A deep-learning (DL) algorithm, using only confocal scanning laser ophthalmoscopy (cSLO) imaging, can predict visual impairment in patients with retinitis pigmentosa (RP), according to a study published in the British Journal of Ophthalmology. 

“The efficiency of clinical trials for RP treatment is limited by the screening burden and lack of reliable surrogate markers for functional end points,” the researchers explain. “Automated methods to determine visual acuity (VA) may help address these challenges.” 

Researchers sought to determine the feasibility of creating a cSLO-based model to predict visual impairment in patients with RP using DL. They retrospectively obtained Snellen corrected VA and cSLO imaging. 

The Johns Hopkins University (JHU) dataset was used for 10-fold cross-validations and internal testing, and included patients with RP with and without molecular confirmation. The Amsterdam University Medical Centers (AUMC) dataset was used for external independent testing, and included molecularly confirmed patients with RP. Both datasets excluded visually significant media opacities and images not centered on the central macula. Utilizing transfer learning, 3 versions of the ResNet-152 neural network were trained: infrared (IR), optical coherence tomography (OCT) and combined image (CI). 

The JHU database included 2569 images from 462 eyes of 231 patients (47% men). In this group, the median age at the time of imaging was 52 years (range: 7-88 years) and the median Snellen VA was 20/40 (range: 20/16 to no light perception). In internal testing (JHU dataset), the area under the curve (AUC) for discriminating between Snellen VA 20/40 or better and worse than Snellen VA 20/40 was 0.83, 0.87 and 0.85 for IR, OCT and CI, respectively.

The AUMC database included 349 images from 166 eyes of 83 patients (60% men). In this group, the median age at the time of imaging was 38 years (range: 6-77 years) and the median Snellen VA was 20/32 (range: 20/16 to light perception). In external testing (AUMC dataset), the AUC for discriminating between Snellen VA 20/40 or better and worse than Snellen VA 20/40 was 0.78, 0.87 and 0.85 for IR, OCT and CI, respectively.

“A DL algorithm that can accurately predict VA from OCT images could be valuable in terms of clinical care and clinical trials for RP” according to researchers. “With further optimization, DL-based OCT analysis could be potentially developed into an OCT ‘potential acuity meter’ and circumvent the impact of undercorrected refractive error, ocular surface disease, and media opacity.”

Study limitations included its retrospective design, the performance of the algorithm decreased in the presence of additional structural abnormalities except for outer retinal atrophy, and at this phase of development, the algorithm can only perform binary classifications rather than multivalent classifications.

Reference

Liu TYA, Ling C, Hahn L, Jones CK, Boon CJF, Singh MS. Prediction of visual impairment in retinitis pigmentosa using deep learning and multimodal fundus images. Br J Ophthalmol. Published online July 27, 2022. doi:10.1136/bjo-2021-320897