A Dual Modality Improves Dx of Glaucomatous Optic Neuropathy
By Lynda Seminara
Selected by Stephen D. McLeod, MD
Journal Highlights
Ophthalmology, February 2022
Download PDF
Xiong et al. developed a bimodal algorithm that combines visual field (VF) data and peripapillary circular OCT to detect glaucomatous optic neuropathy (GON). They compared the accuracy of their dual algorithm to that of each of its components and found that the dual modality outperformed either VF or OCT alone.
For this study, the authors gathered 2,463 pairs of VF and OCT images representing eyes of 1,083 Chinese patients. The majority of image pairs were used to train the FusionNet artificial intelligence algorithm to detect GON, which was defined as retinal nerve fiber layer thinning with corresponding VF defects. The other images were used for testing and validation. The dual algorithm incorporates pattern deviation probability plots from VF reports and from circular peripapillary OCT scans. VF data were collected with the Humphrey field analyzer, and OCT images were obtained from three devices.
The pairs of VF and OCT images were grouped into four datasets, with 1,567 pairs used for training, 441 for primary validation, 255 for internal testing, and 200 for external testing. In addition, four glaucoma specialists classified the cases independently. The main outcome measure was the diagnostic performance of FusionNet versus algorithms based only on visual field data (VFNet) or OCT data (OCTNet).
In the primary validation set, area under the receiver operating characteristic curve (AUC) was .950 for FusionNet, .868 for VFNet, and .809 for OCTNet. FusionNet also outperformed two glaucoma specialists (AUC, .882 and .883). In the internal and external test sets, the combination algorithm performed better than VFNet or OCTNet. Internal AUC data were .917, .854, and .811, respectively. The corresponding external AUCs were .873, .772, and .785. There was only one significant difference between human assessment and FusionNet among the internal and external sets: the AUC for one specialist was .858. Generalizability was good across the three types of OCT.
These results support the hypothesis that multimodal machine-learning models are effective for GON detection. Multicenter studies are warranted to validate FusionNet in different ethnicities and with data from primary eye care providers, said the authors.
The original article can be found here.