Purpose: To develop deep learning (DL) algorithms to detect highly myopic glaucoma using macular microvasculature assessed with swept–source optical coherence tomography angiography (SS–OCTA), and to differentiate glaucomatous optic neuropathy (GO...
Purpose: To develop deep learning (DL) algorithms to detect highly myopic glaucoma using macular microvasculature assessed with swept–source optical coherence tomography angiography (SS–OCTA), and to differentiate glaucomatous optic neuropathy (GON) from non-glaucomatous optic neuropathy (NGON) using retinal nerve fiber layer (RNFL) and optic disc photographs.
Methods: This study included a total of 260 pairs of macular OCTA and OCT images from 260 eyes, comprising 203 eyes with highly myopic glaucoma and 57 eyes with healthy high myopia. In addition, a total of 765 pairs of RNFL and optic disc images from 765 eyes, including 618 eyes with GON and 147 eyes with NGON, were utilized. The DL models were trained, validated and tested using the macular OCTA and OCT images to differentiate highly myopic glaucoma from healthy high myopia, and the RNFL and optic disc photographs to discriminate GON from NGON. The main outcome measure was the area under the receiver operating characteristic curve (AUC) of the DL models.
Results: For differentiating highly myopic glaucoma from healthy high myopia, the DL model achieved an AUC of 0.946 with the OCTA superficial capillary plexus (SCP) images, which was comparable to that with the OCT GCL+ (AUC of 0.982, P = 0.268) or OCT GCL++ images (AUC of 0.997, P = 0.101), and significantly superior to that with the OCTA deep capillary plexus (DCP) images (AUC of 0.779, P = 0.028). For discriminating GON from NGON, the DL model revealed an AUC of 0.98 with the RNFL images, which was comparable to that with the optic disc (AUC of 0.99, P = 0.23) or combined RNFL and optic disc images (AUC of 0.97, P = 0.67), and was significantly superior to that with the masked RNFL (AUC of 0.94, P < 0.05) or combined masked RNFL and optic disc images (AUC of 0.96, P < 0.05).
Conclusion: The DL models, with various imaging modalities including macular OCTA SCP images as well as RNFL and optic disc images, demonstrated excellent performance in detecting glaucoma, suggesting the potential value of our DL models in clinical practice by helping clinicians make accurate diagnoses and treatment decisions.