Automated quantification of CNV volume using deep learning AI algorithm

Abstract

Purpose : An artificial intelligence (AI) algorithm was applied in combination with spectral-domain optical coherence tomography (SD-OCT) to quantify the volume of choroidal neovascularization (CNV) lesions in the mouse CNV model.

Methods : Choroidal neovascularization and vascular leakage were induced using a 532 nm laser in the mouse. The development of CNV lesions were monitored in vivo by fluorescent angiography (FA) and SD-OCT at days 5, 10 and 14 post-induction. The efficacy of systemically administered aflibercept (Eylea®, Bayer AG, Germany) at doses of 5, 15 or 25 mg/kg was evaluated.
The analysis of the CNV leakage area was performed by manually measuring the leakage area from FA images. This analysis was benchmarked against a convolutional neural network (CNN) trained to recognize and quantify the CNV lesion using Aiforia® Cloud (Aiforia Technologies Ltd, Finland) using semantic segmentation and supervised learning. After follow-up period choroidal flat mounts were stained by an endothelial marker, isolectin GS-IB₄.

Results : A significant dose-dependent decrease in the CNV lesion development was seen in aflibercept-treated animals as compared to controls. The 25 mg/kg dose significantly decreased CNV leakage area by 42% (p = 0.046) at day 5. Both 5 mg/kg and 15 mg/kg aflibercept-treated mice did not show statistically significant decrease in CNV leak as compared to control mice. The AI-based methodology for lesion volume analysis (CNN model) at Day 14 showed that groups treated with aflibercept at a dose of 15 mg/kg or 25 mg/kg decrease CNV volume by an average of 53%. Moreover, a decrease in isolectin-positive areas was found in the choroidal flat mounts stainings from the 25 mg/kg and 15 mg/kg dose groups as compared to saline treated mice.

Conclusions : The automated lesion volume analysis and the area of vascular leakage measured with FA showed similar results indicating that systemically administered aflibercept presents a dose-dependent reduction on CNV lesion development. The use of AI-driven CNV volume evaluation represent an unbiased and standardized methodology for CNV pathology quantification.

This is a 2020 ARVO Annual Meeting abstract.