Using AI to Differentiate Glaucomatous and Compressive Optic Neuropathy
By Lynda Seminara
Selected By: Prem S. Subramanian, MD, PhD
Journal Highlights
British Journal of Ophthalmology
Published online Feb. 25, 2020
Download PDF
Lee et al. evaluated whether a novel deep learning (DL) classifier could discriminate between glaucomatous optic neuropathy (GON) and compressive optic neuropathy (CON). They found that their transfer learning‒trained model accurately distinguished GON from CON and outperformed clinical diagnostic parameters.
The researchers’ DL model uses ganglion cell‒inner plexiform layer (GCIPL) and retinal nerve fiber layer (RNFL) maps obtained via spectral-domain optical coherence tomography (SD-OCT). For this study, they recruited their study population from a Korean database. Bottleneck features from four images were integrated and used as training data for the classifier’s deep neural network. Area under the curve (AUC) was calculated to validate and compare the performance of the DL classifier with that of standard diagnostic parameters, such as SD-OCT thickness profiles.
Overall, 80 patients with GON and 54 patients with CON were included, along with 80 and 81 SD-OCT image sets, respectively. Baseline characteristics were similar for the study groups. When discriminating GON from CON, the DL classifier achieved AUC of 0.990, sensitivity of 97.9%, and specificity of 92.6%. Moreover, it significantly outperformed conventional diagnostic parameters: temporal raphe sign (AUC, 0.804), superonasal GCIPL thickness (AUC, 0.815), and superior GCIPL thickness (AUC, 0.776); all p < .001. Assessment of heat maps—derived from RNFL deviation maps that highlighted OCT areas for which the DL algorithm was the best predictor—indicated that the model used clinically important information to interpret the images.
According to the authors, the DL classifier’s discrimination of GON from CON, even if the clinical diagnosis was unclear, supports its potential to augment diagnostic accuracy and objectivity in ophthalmology. With additional training and a larger dataset, the DL model could be helpful when visual field defect patterns are equivocal.
The original article can be found here.