Automated Regional Detection of Intervortex Venous Anastomosis on Ultra-Widefield Indocyanine Green Angiography Using a Two-Stage Deep Learning Framework

Objectives: This study aimed to develop and validate a deep learning–based system for automated detection of intervortex venous anastomosis (IVA) on ultra-widefield indocyanine green angiography (UWF-ICGA). Methods: A total of 183 UWF-ICGA images with an original resolution of 3900 × 3072 pixels were included. A two-stage deep learning framework was developed. In the first stage, an object detection model automatically localized four predefined quadrants within each UWF-ICGA image. In the second stage, each extracted quadrant was classified according to the presence or absence of IVA. Quadrants demonstrating definite venous connections between adjacent vortex vein drainage territories were defined as IVA-positive, whereas quadrants without identifiable intervortex venous connections were defined as IVA-negative. Brightness normalization was applied before analysis, and images were resized to 512 × 512 pixels using linear interpolation. The object detection model was trained using 155 images and tested using 28 images. The 183 original images generated 732 quadrant images, of which 622 were used for training and 110 for testing of the classification model. Model performance was evaluated using accuracy, precision, recall, F1 score, and confusion matrices. Results: The object detection model correctly localized all predefined quadrants in the test dataset, achieving an accuracy, precision, recall, and F1 score of 100.0%. All 112 manually labeled regions from 28 test images were correctly matched by the model. The IVA classification model achieved an accuracy of 93.63%, precision of 95.39%, recall of 91.46%, and F1 score of 93.38%. Among 110 test quadrants, all 69 IVA-negative quadrants were correctly classified, while 34 of 41 IVA-positive quadrants were correctly identified. Seven IVA-positive quadrants were misclassified as negative, whereas no IVA-negative quadrant was classified as positive. When IVA-positive was considered the positive class, the corresponding sensitivity and specificity were 82.9% and 100%, respectively. Conclusions: A two-stage deep learning system enabled automated detection of IVA on UWF-ICGA with high classification performance. Automated assessment of IVA may provide an objective and reproducible approach for evaluating choroidal venous remodeling and may serve as a platform for future quantitative investigations of vortex vein abnormalities.