Cornea topography maps allow ophthalmologists to screen and diagnose cornea pathologies. We aim to automatically identify any cornea abnormalities based on such cornea topography maps, with focus on diagnosing keratoconus. To do so, we represent the OCT scans as images and apply Convolutional Neural Networks (CNNs) for the automatic analysis. The model is based on a state-of-the-art ConvNeXt CNN architecture with weights fine-tuned for the given specific application using the cornea scans dataset. A set of 1940 consecutive screening scans from the Saarland University Hospital Clinic for Ophthalmology was annotated and used for model training and validation. All scans were recorded with a CASIA2 anterior segment Optical Coherence Tomography (OCT) scanner. The proposed model achieves a sensitivity of 98.46% and a specificity of 91.96% when distinguishing between healthy and pathological corneas. Our approach enables the screening of cornea pathologies and the classification of common pathologies like keratoconus. Furthermore, the approach is independent of the topography scanner and enables the visualization of those scan regions which drive the model’s decisions.

Cornea topography maps allow ophthalmologists to screen and diagnose cornea pathologies. We aimto automatically identify any cornea abnormalitiesbased on such cornea topography maps, with focus on diagnosing keratoconus. A set of 1946 consecutive screening scans from the Saarland University Hospital Clinic for Ophthalmology was annotated and used for model training and validation. All scans were recorded witha CASIA2 anterior segment Optical Coherence Tomography (OCT) scanner.We propose to represent the OCT scans as images and apply Convolutional Neural Networks (CNNs) for the automatic analysis. The developed model is based on a state-of-the-art ConvNeXt CNN architecture with weights fine-tuned for the given specific application using the cornea scans dataset. On a new dataset, our model achieves a sensitivity of 97% and a specificity of 97% when distinguishing between Healthy and Pathological corneas. While a comparison to previous work is intricate due tosignificant variations in the experimental setup, our model outperforms other published studies, either in terms of detection performance, and/or in terms of number of potential cornea abnormalities the model can identify. Furthermore, the proposed approach is independent of the topography scanner and allows to visually represent scan regions that drive the models' decision.