Validation of a novel artificial intelligence (AI) machine learning algorithm for quantification of vascular pathology in preclinical models

Abstract

Purpose : To validate a novel artificial intelligence (AI) machine learning algorithm for the automated quantification of lesion size in the mouse choroidal neovascularization (CNV) model.

Methods : CNV was induced in male C57Bl/6JRj mice using a diode laser. The progression of CNV was monitored using live in vivo imaging: fluorescent angiography and spectral-domain optical coherence tomography (SD-OCT). SD-OCT scans covering the whole retina were analyzed by a proprietary algorithm, which uses a combination of convolutional neural network (CNN) and traditional computer vision algorithms. The neural network was trained to recognize and quantify CNV lesions using a transfer learning approach.

Results : We used the Dice coefficient to evaluate AI algorithm-derived results. Dice coefficient comparing predicted lesion volume with volume from the manually annotated data was 0.87 and 0.89 for follow-up day 3 and day 7, correspondingly.

Conclusions : Our data provide evidence that our novel AI algorithm can successfully detect and quantify CNV lesions in the mouse CNV model. Notably, the algorithm accurately recognized temporal changes in retinal structure and pathology over a one-week experimental period. Automated vs. manual analysis resulted in approx. 90% congruence. The novel AI algorithm will provide automated computational analysis of phenotypes in the commonly-used CNV model for age-related macular degeneration, and offer complementary quantitative and bias-free data to accelerate and support drug discovery efforts.

This is a 2021 ARVO Annual Meeting abstract.