Abstract
Purpose :
To develop a deep learning construct to detect Uveitis from retinal fundus photographs of Experimental Autoimmune Uveitis (EAU) mice and to propose a framework for consistent and reproducible characterization of images related to animal research studies.
Methods :
Accuracy, consistency, and reproducibility are critical for animal model characterization. We developed a deep learning model (DLUveitis) based on the VGG16 architecture to grade uveitis from retinal fundus photographs of EAU mice (Fig. 1). We utilized 1800 fundus photographs to train and test the models for detecting five levels of disease severity and used 300 images to independently validate the findings. We further developed a method to calculate the detailed clinical score of EAU and visualize the outcome of different models (Fig. 2)
Results :
We analyzed 2100 fundus photographs that corresponded to normal, trace, moderate, advanced and severe stages of the uveitis disease. Three human expert readers annotated the training images based on evaluations from both fundus photographs and optical coherence tomography (OCT) images. Based on the initial dataset of 1800 fundus images, the AUC of the model in distinguishing five levels of severity was 0.98 (95%CI, 0.97-0.99). Based on the additional independent validation subset of 300 images, the AUC of the model in distinguishing five levels of severity was 0.96 (95%CI, 0.93-0.99). Our model outperforms human graders.
Conclusions :
Animal models are invaluable tools for studying human diseases and drug testing. The proposed deep learning construct accurately detects uveitis and five severity levels from fundus photographs of EAU mice. In addition, the clinical score generated by DLUveitis provides more details about the disease severity, compared to human experts, thus providing a highly consistent and reproducible pilot model for subsequent animal research studies.
This is a 2021 ARVO Annual Meeting abstract.