AI May Help Identify Candidates for Refractive Surgery
By Lynda Seminara
Selected and Reviewed By: Neil M. Bressler, MD, and Deputy Editors
Journal Highlights
JAMA Ophthalmology, May 2020
Download PDF
Xie et al. evaluated the utility of deep learning as an adjunct to tomographic imaging for identifying high-risk corneas. Their artificial intelligence (AI) system appeared useful for classifying images, providing important details about the cornea, and identifying potentially at-risk corneas.
This cross-sectional analysis was performed at the Zhongshan Ophthalmic Center in Guangzhou, China. The researchers included patients throughout China who wanted refractive surgery, had a primary diagnosis of keratoconus, and had a stable post-op refractive state. Data were collected using a Pentacam HR system, and four-map composite refractive images were used to determine the overall profile of the cornea. Altogether, data for 6,465 de-identified corneal tomographic images (1,385 patients) were used to generate the AI model, which was based on the Pentacam InceptionResNetV2 Screening System (PIRSS). Images were analyzed independently by 20 individuals (including 10 senior ophthalmologists) and by the AI model.
The overall accuracy rate of the PIRSS model was 94.7% (95% confidence interval [CI], 93.3%-95.8%) on the validation dataset. Most areas under the receiver operator characteristic curves were above 0.99. For an independent test dataset, the model achieved similar accuracy (95% [95% CI, 88.8%-97.8%]), comparable to that of five senior ophthalmologists who perform refractive surgery (92.8% [95% CI, 91.2%-94.4%]; p = .72). The PIRSS model was superior to human classifiers for identifying corneas that would be unsuitable for refractive surgery (95% vs. 81%; p < .001).
Larger samples and other refinements are needed to improve the performance of PIRSS, said the authors, who emphasized that technology cannot replace human clinical expertise. They suggested that biochemical assessment may improve screening for keratoconus and that combining it with AI could help guide clinical decisions. (Also see related commentary by Travis K. Redd, MD, MPH, J. Peter Campbell, MD, MPH, and Michael F. Chiang, MD, MA, in the same issue.)
The original article can be found here.