Deep Learning for Automated Assessment of Retinal Degeneration in Fundus Autofluorescence Images

Abstract

Purpose : Fundus autofluorescence (FAF) imaging is a common method to assess and monitor the progression of retinal degeneration in in vivo studies. The conventional manual assessment of degenerated areas in FAF images is known for its time-intensive nature and inherent challenges, prompting the need for an accurate automated solution. Our proposed method employs deep-learning techniques to automate the quantification of degenerated regions in FAF images, facilitating a precise analysis of retinal degeneration.

Methods : A Unet deep learning model was developed using FAF image data from the Royal College of Surgeons (RCS) rats, which is a common animal model for inherited retinal dystrophy. FAF images were acquired using cSLO (SPECTRALIS HRA+OCT, Heidelberg, Germany). Baseline imaging was obtained to rule out rats with pre-existing defects in the retina or optic nerve. After the establishment of the disease in adult RCS rats, the captured image data was partitioned into 80% training data and 20% testing data. Retinal degeneration areas were freehand-selected and quantified using ImageJ (NIH, USA). For our analysis, the optic nerve and blood vessels were excluded from the area of interest. A correlation study was employed to assess the performance of our methodology.

Results : Our Unet deep learning model was able to effectively identify regions exhibiting retinal degeneration in FAF images of RCS rats. The mean ± SD percentage of the retinal degeneration area using our deep learning model and the manual method was 9.23% ± 3.23% and 9.34% ± 2.15%, respectively. The correlation analysis between the two methodologies yielded a Pearson correlation coefficient (r) of 0.81.

Conclusions : Through the deep learning model, we could accurately quantify the area of retinal degeneration in FAF images, eliminating the need for manual selection and mitigating potential user bias. Further optimization of performance is anticipated through the expansion of the database size.

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