Deep Learning and Radiomics Assessment for Highly Myopic Glaucoma Detection Based on Fundus Photography

To investigate spatially distinctive features in fundus photographs of highly myopic glaucoma (HMG) by integrating radiomics and deep learning.

Cross-sectional study.

Semi-automated optic disc segmentation was performed on 2000 images sourced from the Retinal Fundus Glaucoma Challenge Edition and Pathologic Myopia Challenge public data sets. We trained models with 628 images, including 217 cases of high myopia (HM), 221 cases of primary open-angle glaucoma (POAG), and 190 cases of HMG. An external validation set of 106 images was collected including 31, 36, and 39 fundus photographs from patients with HM, HMG, and POAG, respectively.

Semi-automated optic disc segmentation was performed by U-Net combined with manual delineation. Additionally, 5 regions of interest (ROIs) covering the optic disc and its surrounding region were explored. Fundus photography-based radiomics feature selection and optimization were conducted using random forest and support vector machine algorithms, which underwent fivefold cross-validation. Model performance was evaluated for radiomics, clinical, and combined models. An external validation set was used to evaluate the models performance, and we also examined radiomics features varied across glaucoma stages.

A total of 414 radiomics features were extracted from 5 regions of interest. We evaluated the model performance using accuracy, recall, precision, F1 score, and area under the receiver operating characteristic curve (AUROC). The Youden index was used to determine the cut-off value, and sensitivity and specificity were calculated.

The U-Net-based optic disc segmentation reached a Dice Similarity Coefficient of 0.95. The radiomics model detected HMG from HM and POAG achieved the accuracies of 0.97 and 0.85, respectively. Both results exceeded the performance of the clinical model, which achieved 0.90 and 0.71. The top radiomics features achieved AUROC of 0.984 and 0.855 for HMG versus HM and HMG versus POAG in the external validation, respectively, including intensity features in the extra-optic nerve and textural features in the optic nerve. These features showed strong diagnostic capability when stratified by Youden values and exhibited independence from glaucoma progression.

The study combined U-Net and radiomics to delineate the spatial distribution biomarkers of HMG, establishing a quantitative classification model correlated with anatomical features.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.