Related Experiment Video
Updated: Jun 9, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Automated Triage for New Keratoconus Referrals Using Multimodal Deep Learning
Shafi Balal1,2, Lynn Kandakji2, Marcello Leucci1
1Moorfields Eye Hospital NHS Trust, London, UK.
Purpose:
To develop and validate deep learning models for predicting keratoconus progression risk using multimodal imaging and clinical data, enabling risk-stratified patient monitoring.
Design:
A retrospective cohort study with external validation.
Participants:
We analyzed 7396 eyes of 3893 patients (internal dataset) and 963 eyes of 519 patients (external validation dataset) with keratoconus who underwent MS-39 (CSO Italia) anterior-segment OCT (AS-OCT) and Placido topography between October 2020 and June 2024.
Methods:
Progression was defined using the global consensus criteria requiring changes in multiple parameters above device-specific precision limits. We compared conventional machine learning, unimodal deep learning, and multimodal fusion architectures to predict 2-year progression risk from baseline data. We developed recurrent neural networks to incorporate data from sequential clinic visits. We assessed clinical utility through simulated risk-stratified triage pathways in an external validation cohort.
Main Outcome Measures:
The area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and predictive values for keratoconus progression within 2 years were measured.
Results:
The multimodal model of AS-OCT, Placido, and tabular data yielded the best single-visit prediction (AUROC 0.84, 95% confidence interval [CI]: 0.83-0.85); adding sequential visits via long short-term memory boosted this to 0.93 (95% CI: 0.91-0.96). Fifty-eight percent of patients could be classified as low risk (90% chance of correctly predicting no progression in 2 years) after their baseline clinic visit, rising to 83% with inclusion of data from a second clinic visit.
Conclusions:
Artificial intelligence models can be used to develop risk-stratified pathways for monitoring keratoconus, reducing unnecessary follow-up for patients at low risk of progression while ensuring optimal resource allocation and timely intervention for those at high risk.
Financial Disclosures:
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
