The study utilized macular-OCT data from the UK Biobank dataset, which included spectral domain macular-OCT images from over 87,000 participants. Glaucoma cases and healthy individuals were selected from this dataset, with specific criteria for inclusion and exclusion. A total of 255 individuals with glaucoma and 2,812 healthy participants were included in the study. The dataset was split into training, validation, and test sets using a tenfold cross-validation strategy.
Data preprocessing steps were applied to B-scans and OCT volumes before training the models. A 2D model using ResNet18 architecture was trained on B-scans to predict glaucoma, with XGBoost used to aggregate the predictions. For the 3D models, adaptations of ResNet18 and DenseNet121 were explored, along with a 3D-CNN-Encoder model. Transfer learning was utilized to adapt pretrained weights from 2D models to the 3D counterparts. The models were fine-tuned with specific hyperparameters and trained for 50 epochs.
Performance metrics such as accuracy, sensitivity, specificity, and AUC-ROC were used to evaluate the models. The Youden Index was employed to find the best balance between sensitivity and specificity for each model and dataset. The study also utilized XGBoost for aggregating results from OCT B-scan predictions and reported on the age and sex distribution of the study sample. Overall, the study focused on developing and evaluating deep learning models for glaucoma detection using OCT data.
Source link
Source link: https://www.nature.com/articles/s41598-024-62411-6