AI Quantification of Vascular Lesions in Mouse Fundus Fluorescein Angiography

Abstract

Purpose: Quantifying vascular leakage in fundus fluorescein angiography (FFA) is a critical endpoint in preclinical models of diseases such as neovascular age-related macular degeneration, retinopathy of prematurity, and diabetic retinopathy. Traditional manual methods are labor intensive and prone to variability. We developed an artificial intelligence (AI)-assisted method to improve efficiency and accuracy in quantifying vascular lesions in FFA images.

Methods: Nikon NIS-Elements software with AI functionality was used to create an automated FFA analysis method. FFA images were acquired using the Phoenix MICRON IV imaging system in two mouse models of ocular angiogenesis: (1) very low-density lipoprotein receptor (Vldlr) knockout mice exhibiting spontaneous pathological chorioretinal neovascularization, and (2) a laser-induced choroidal neovascularization model. The AI model was trained on manually segmented FFA images to delineate lesions and quantify lesion area and fluorescence intensity.

Results: The AI model demonstrated high accuracy in quantifying vascular lesions in FFA images, achieving 99.7% agreement with manual counts. It attained a precision, recall, and F1 score of 0.94, with an intraclass correlation coefficient (ICC) of 0.991. The model showed strong spatial agreement with manual segmentations and consistent lesion area measurements. On validation images, it maintained expert-level performance (ICC = 0.998) with high sensitivity and precision. Additionally, it effectively captured temporal changes in vascular leakage by measuring lesion area and fluorescence intensity, demonstrating robustness in real-world experiments.

Conclusions: Our AI model quantifies vascular lesions in FFA images with high accuracy, outperforming manual analysis.

Translational relevance: AI-based quantification provides a scalable, consistent alternative to manual methods, enhancing research efficiency.

Figure 1.

Evaluation of the AI method performance in lesion detection on training images. (A) Heatmap showing the lesion count in each of the 70 training images. (B) Venn diagram illustrating lesion detection overlap between AI and manual methods. (C) Correlation analysis of lesion counts per image between the ground truth and AI methods, quantified by the ICC. (D) Distribution of percentage deviation in AI-predicted lesion counts per image compared to the ground truth, represented using a box-and-whisker plot and a pie chart. (E) Bland–Altman plot comparing lesion counts per image between the ground truth and AI methods. (F) Performance metrics, including precision, recall, and F1 score, for lesion detection by the AI method.

Figure 2.

Evaluation of the AI method performance in lesion spatial agreement and area on training images. (A) Spatial agreement at the level of individual lesion between ground truth and AI method quantified by IoU. (B) Spatial agreement at the level of individual image between ground truth and AI method quantified by IoU. (C) Correlation analysis of lesion area per image between the ground truth and AI methods, quantified by the ICC. (D) Bland–Altman plot comparing lesion area per image between the ground truth and AI methods. (E) Distribution of percentage error in AI-predicted lesion area per image compared to the ground truth, represented using a box-and-whisker plot and frequency table.

Figure 3.

Evaluation of the AI method performance in individual lesion area measurement on training images. (A) Correlation analysis of individual lesion area measurement between the ground truth and AI methods, quantified by the ICC. (B) Bland–Altman plot comparing individual lesion area measurement between the ground truth and AI methods. (C) Distribution of percentage error in AI-predicted individual lesion area measurements compared to the ground truth, represented using a box-and-whisker plot and frequency table.

Figure 4.

Evaluation of the AI method performance in lesion detection and spatial agreement on validation images. (A) Heatmap showing the lesion count in each of the nine validation images. (B) Venn diagram illustrating lesion detection overlap between AI and manual methods. (C) Correlation analysis of lesion counts per image between the ground truth and AI methods, quantified by the ICC. (D) Bland–Altman plot comparing lesion counts per image between the ground truth and AI methods. (E) Performance metrics, including precision, recall, and F1 score, for lesion detection by the AI method. (F) Spatial agreement at the level of individual lesion between ground truth and AI method quantified by IoU. (G) Spatial agreement at the level of individual image between ground truth and the AI method quantified by IoU.

Figure 5.

Evaluation of the AI method performance in lesion area measurement on validation images. (A) Correlation analysis of individual lesion area measurement between the ground truth and AI methods, quantified by the ICC. (B) Correlation analysis of image-wise cumulative lesion area measurement between the ground truth and AI methods, quantified by the ICC. (C) Bland–Altman plot comparing individual lesion area measurement between the ground truth and AI methods. (D) Bland–Altman plot comparing image-wise cumulative lesion area measurement between the ground truth and AI methods. (E) Distribution of percentage error in AI-predicted individual lesion area measurements compared to the ground truth, represented using a box-and-whisker plot. (F) Distribution of percentage error in AI-predicted image-wise cumulative lesion area measurements compared to the ground truth, represented using a box-and-whisker plot.

Figure 6.

Evaluation of the AI method performance in measuring vascular leakage. (A) Manual and AI-assisted assessment of vascular leakage as measured by expansion of lesion area. Areas of individual lesions are plotted. P < 0.05 (two-tailed unpaired t-test) indicates significance. (B) Manual and AI-assisted assessment of vascular leakage as measured by expansion of lesion area. Cumulative lesion areas of individual images are plotted. P < 0.05 (two-tailed paired t-test) indicates significance. (C) Manual and AI-assisted assessment of vascular leakage by measuring the mean fluorescence intensity of individual lesions. The fluorescence intensities of individual lesions are plotted. P < 0.05 (two-tailed unpaired t-test) indicates significance. (D) Manual and AI-assisted assessment of vascular leakage. Mean fluorescence intensities across all lesions within each image are plotted. P < 0.05 (two-tailed paired t-test) indicates significance.

Figure 7.

Evaluation of the AI model performance on peripheral lesion detection and quantification. (A) Representative FFA images acquired at 5 and 9 minutes, highlighting the ability of the AI model to accurately segment lesions in peripheral regions where the optic nerve head is not centered are presented. (B) AI-assisted assessment of vascular leakage in peripheral lesions as measured by expansion of lesion area. Areas of individual lesions are plotted. P < 0.05 (two-tailed unpaired t-test) indicates significance. (C) AI-assisted assessment of vascular leakage in peripheral lesions as measured by expansion of lesion area. Cumulative lesion areas of individual images are plotted. P < 0.05 (two-tailed paired t-test) indicates significance. (D) AI-assisted assessment of vascular leakage in peripheral lesions, by measuring the mean fluorescence intensity of individual lesions. The fluorescence intensities of individual lesions are plotted. P < 0.05 (two-tailed unpaired t-test) indicates significance. (E) AI-assisted assessment of vascular leakage in peripheral lesions, by measuring the mean fluorescence intensity across all lesions within each image are plotted. P < 0.05 (two-tailed paired t-test) indicates significance.