Development of an unsupervised deep learning Keratoconus classification model through analysis of Keratoconus patients from multiple centres

Abstract

Purpose : One of the greatest challenges of Keratoconus (KC) is the lack of consistency and agreement on an accurate classification system, leading to significant issues predicting individual patient progression. We sought to develop an unsupervised deep learning convolutional neural network (CNN) model to accurately classify KC patients into 4 classes. We applied this model to retrospectively collected datasets from clinics in Australia and Saudi Arabia to compare results across differentiated datasets, comparing accuracy to the Amsler-Krumeich (AK) classification model.

Methods : The study utilised 285 KC patients, 243 from Australia and 42 from centres in Saudi Arabia. We developed an unsupervised CNN comprising a Multilayer Perceptron for data classification, and a Variational Autoencoder utilising Gaussian sampling for clustering. We collected 27 variables for each patient, a combination of Pentacam (Oculus) output and clinical information, which is cleaned into a tabular format and applied to the model. Our model is significantly differentiated from other studies with machine learning and KC, as we apply the model to tabular rather than image data. Data was applied to the deep learning CNN for unsupervised classification, and then compared against the AK system as the ground truth.

Results : The model results in 83.6%-86.2% accuracy at independently classifying Keratoconus patients against AK, and receiver operating characteristics 81%-96% (Table 1) demonstrating its performance at differentiating between the 4 classes. Figure 1 graphs our actual patient AK classifications determined by clinicians. Figure 2 graphs our unsupervised model’s independent classification of the same patients, demonstrating its accuracy.

Conclusions : Accurate classification of KC patients is vital to improve vision outcomes. Our CNN model results in accuracy at 86.2% classifying patients into 4 classes, compared to the AK system. The model demonstrates great potential for application to broader data sets, which will improve performance.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Table 1 Results

 

Figure 1 and 2