Use of deep learning-based algorithms to detect keratoconus changes in the color-coded corneal maps

Background and aims: The use of artificial intelligence techniques developed from deep learning(DL) based algorithms has increased significantly in recent years. Early and accurate detectionof keratoconus provides opportunities to address risk factors and offer treatments to potentiallyslow its progression. One of the suitable neural networks for recognizing and classifyingimages in keratoconus is the convolutional neural networks (CNN) technique. For this purpose,the present study aimed to develop a CNN model for the automatic detection of keratoconus usingstandard color-coded corneal maps (curvature, elevation, and pachymetry maps) obtained by theScheimpflug corneal imaging technique.Method: This multicenter retrospective study included corneal maps of keratoconic and healthysubjects. The corneal tomographic maps considered for each scan were the axial curvature map,the anterior and posterior elevation map, and the pachymetry map obtained using the PentacamHR (Oculus Optikgeräte, Wetzlar, Germany). Keratoconus eyes were classified into four stagesbased on the Amsler-Krumeich grading system. All scans were categorized into five differentclassification tasks: healthy versus keratoconus, healthy versus keratoconus stage ۱, keratoconusstage ۱ versus ۲, keratoconus stage ۲ versus ۳, and ۵-class classification between healthy and eachstage of keratoconus. Considering that each of the four maps contained numerous parameters todetect keratoconus, four CNN models were trained with one for each corneal map. A fifth modelwas added which used a concatenation of four corneal maps. Various classification models wereimplemented using Python V.۳.۷, Keras V.۲.۳.۱ and Tensorflow V.۱.۱۴ as backend, and all modelswere trained by using the Adam optimizer. Accuracy, sensitivity, specificity and area under thereceiver operating characteristic curve were used to assess the diagnostic ability of each model.Results: ۱۹۲۶ corneal tomography scans including ۱۷۰۲ patients with keratoconus and ۱۳۴healthy controls were assessed. A CNN model detected keratoconus versus normal eyes withan accuracy of ۰.۹۷۸۵ (۹۵% CI: ۰.۹۶۴۲ to ۰.۹۹۲۸), considering all four maps concatenated. Theaccuracy of the CNN models considering the axial curvature map, pachymetry map, anterior andposterior elevation maps were independently ۰.۹۲۸۳, ۰.۹۶۴۲, ۰.۹۶۴۲, and ۰.۹۷۴۹, respectively.The accuracy of concatenated CNN models in differentiating between healthy corneas and keratoconusstage ۱ was ۰.۹۰, between keratoconus stages ۱ versus ۲, keratoconus stages ۲ versus ۳was ۰.۹۰۳۲, and ۰.۸۵۳۷, respectively.Conclusion: Accurate automated detection of keratoconus and its evolution is possible using aCNN based on four color-coded corneal map images obtained by the Scheimflug technique. CNNmodels provide excellent performance for the detection and staging of keratoconus in a clinicalsetting, as precise detection of keratoconus at early stages is still challenging in daily practice.

کلیدواژه ها:

keratoconus ، artificial intelligence ، deep learning ، corneal tomography