Prediction Of Keratoconus Progression Using Multi-Modal Deep Learning

Keratoconus is a common cause of visual disability in young working-age people. This study aimed to determine whether keratoconus progression could be accurately predicted from the first clinic visit using deep learning models trained on baseline clinical data, anterior segment optic coherence tomography (AS-OCT) sectional images, and Placido images of the cornea. Our goal is to use AI assistance to stratify the risk of progression in order to allocate resources for disease monitoring and early intervention effectively. 

The study was conducted at Moorfields Eye Hospital with retrospective data collected between September 2019 and March 2023 using the MS-39 AS-OCT tomography device (CSO, Italy). 

We included patients referred to our keratoconus clinic with multiple visits at least 90 days apart and a minimum of 6-months follow-up. The ground truth label for keratoconus progression was based on the ectasia consensus group criteria (change above the device's measurement error for at least two parameters from curvature and pachymetry). Adaptive thresholds were derived for measurement error, accounting for disease severity. Three artificial intelligence models were developed: XGBoost for tabular and demographic data, a uni-modal convolutional neural network (CNN) for Placido and OCT images, and a multi-modal CNN. Models were trained using PyTorch v3.9 for 50 epochs, with performance validated through k-fold cross-validation.

In total our dataset comprised of 36,673 images (6,684 eyes); mean patient age was 38.3 years (+/-15.2) with 58.3% male. Progression at any time point during follow-up occured in 11.2% of patients within a mean follow-up period of 18.9 months (+/-6.2). The multi-modal CNN achieved the best performance with a ROC-AUC of 0.9. Combined modality approaches of OCT or Placido images with tabular data both yielded ROC-AUC values of 0.84. When evaluating single-modality approaches, Placido topography alone performed best (ROC-AUC = 0.82), followed by OCT imaging (ROC-AUC = 0.79) and tabular data (ROC-AUC = 0.78).

These AI-driven prognostic tools could improve keratoconus patient care by enabling ophthalmologists to identify high-risk patients from their first visit, facilitating earlier intervention or closer monitoring, and ultimately reduce the rate of disease progression. Future work will focus on external validation and prospective implementation studies with local validation to establish their utility in the real-world clinical setting.