Automated machine learning models for detection of neovascularization from ultrawide field directed optical coherence tomography in eyes with advanced diabetic retinopathy

Abstract

Purpose : To develop and validate an automated machine learning model to detect neovascularization elsewhere (NVE) on ultrawide field (UWF) directed cross sectional swept source optical coherence tomography (SS-OCT) scans in eyes with proliferative diabetic retinopathy (PDR).

Methods : Vertex AI Vision (Google Cloud) models were developed using previously acquired UWF-directed 6x6 mm SS-OCT scans from eyes with PDR. Trained graders classified individual scans as to whether they were gradable or ungradable and whether they had NVE or not. Lesions were designated as NVE if retinal vasculature was observed to break through the ILM. The model was developed using a 1,967 image set [(953 (48.4%) images without NVE and 1,014(51.6%) images with NVE] split (8-1-1) between training, validation and testing to detect NVE presence (28 eyes/19 patients). The model was tested using a prospectively acquired data set (12 eyes/8 patients) using the same device [N=639 images 270(42.2%) without NVE and 369(57.7%) with NVE]. Sensitivity and specificity (SN/SP), positive and negative predictive value (PPV/NPV) and accuracy for NV detection were calculated.

Results : Area under the precision-recall curve (AUPRC) was 0.997(figure 1). The model’s overall accuracy was 0.997 with 99% precision and recall. The SN/SP for detecting NVE was 99.0% and 98.9% with a PPV/NPV of 99.0% and 98.9% with an accuracy of 99.0%. The SN/SP on the prospectively acquired dataset for detecting NVE was 0.91/0.98 with a PPV/NPV of 98.5% and 88.33% with an accuracy of 93.7%

Conclusions : These findings suggest that the machine learning models can be a potentially valuable tool for the automated detection of NVE in eyes with PDR, offering high accuracy and reliable identification of these vision threatening features on UWF SS-OCT scans. Integrating such noninvasive automated models into clinical practice might enhance patient care by identifying NVE early. This approach would facilitate timely interventions and management strategies for patients with advanced diabetic retinopathy and possibly allow quantitative NVE measurements not previously possible. Further validation in larger and diverse patient populations is critical to establish the broader applicability and clinical utility of this approach.

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

 

AUPRC showing the performance of the autoML model at detecting NVE on SS-OCT scans