Keratoconus Classification and Segmentation using Deep Learning on Raw Anterior-Segment Optical Coherence Tomorgraphy Imaging

Abstract

Purpose : Machine learning has been applied to keratoconus (KC) detection in recent years with encouraging results. However, most techniques use derived parameters such as topographic heatmaps, rather than analysing raw images with modern deep learning techniques. In this study, KC severity classification is performed using a bespoke convolutional neural network (CNN) applied to raw scans from the MS-39 Anterior Segment Optical Coherence Tomography (AS-OCT) device.

Methods : The dataset consists of AS-OCT volumes from 102 unique eyes from 95 patients that visited the Moorfields Eye Hospital Early Keratoconus Clinic between June 2020 and December 2022. Eyes were split into two groups: 51 stage 1 (K2<48, pachymetry < 520 μm) and 51 stage 4 (K2 >55). 1024x1800 resolution scans at 25 orientations were captured, resulting in a total of 2550 images. The data was then divided into training, test and validation sets using a 70:15:15 split and trained using binary cross-entropy loss and Adam optimizer. Another classification model was trained on segmented scans to allow the network to focus on the cornea rather than other parts of the anterior segment anatomy. The segmentation was obtained by training a U-Net model on 248 manually annotated images equally split between stage 1 and 4. The dataset was split 80:20 for training and test. Patch-based segmentation was used with a patch size of 256x256 and batch size of 16. The network was trained for 40,000 iterations using binary cross entropy loss and Adam optimizer.

Results : The CNN was able to classify disease severity (stage 1 vs stage 4 keratoconus) with a validation accuracy of 95.00% The segmentation model achieved a dice score of 96.92% on the test dataset. However, when applying the masks to the raw images and feeding these to the CNN, the model performance did not improve substantially.

Conclusions : Automatic classification of disease severity may offer diagnostic ability as an automated screening tool, reducing the burden on ophthalmic healthcare professionals. Future research efforts will focus on external validation of the model, the challenge of detecting subclinical keratoconus (stage 0) and combining raw images with other parameters such as biomechanical, demographic and genetic data.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.