Artificial Neural Network For Automated Keratoconus Detection Using A Combined Placido Disc And Anterior Segment Oct Topographer

Deep learning algorithms have proven to be effective in automating keratoconus diagnosis by using placido based topographers. However, new combined placido disc and high-resolution anterior segment OCT (AS-OCT) topographers are currently available, and they provide new valuable data (such as reliable and reproducible corneal epithelial and stromal thicknesses maps) that could feed these neural networks to further enhance their keratoconus recognition capacity.  The aim of the current study is to build and assess the efficacy of an automated program for keratoconus and keratoconus suspect detection based on corneal measurements provided by a combined placido disc and anterior segment OCT topographer (MS-39, CSO).

Vissum Miranza (Alicante, Spain), Magrabi Eye  Hospital (Jeddah, Saudi Arabia), Muscat Eye Laser Center (Muscat, Oman).

Multicentric cross-sectional study. An artificial neural network (ANN) was created using 6677 eyes from equal number of patients (classified as: 2663 normal eyes, 1616 keratoconus eyes, 210 keratoconus suspect eyes, 1519 Myopic post-op eyes, 669 abnormal eyes). Each group was randomly divided into a training set (70% of the dataset) and a validation set (remaining 30%). A multilayer perceptron network with backpropagation learning algorithm was developed for the study. Indexes used to train the ANN were based on curvature and elevation of both the anterior and posterior corneal surfaces, and the new OCT corneal indexes  based on corneal, stromal and epithelial thicknesses.

Our ANN performed very well on identifying keratoconus cases (accuracy of 98.57%, precision of 95.96%, recall of 97.94%, and F1 of 96.94%) and performed well on detecting keratoconus suspect cases (accuracy of 98.46%, precision of 83.64%, recall of 69.70%, and F1 of 76.03%). 

In a previous study, similar methodology was applied using a Scheimpflug with Placido disc topographer (Sirius). With it, by including anterior and posterior corneal surface data, accuracy and precision were 98.2% and 97.9% respectively (results equivalent to our current ANN). However, for subclinical keratoconus detection, accuracy and precision with Sirius were 97.3% and 78.8% respectively. Thus, performing below current ANN including also AS-OCT variables. 

Stablished keratoconus diagnosis has become somehow easy, but its diagnosis on the earliest stages remains challenging, and relies mostly on clinicians experience and intuition. Detecting these suspicious cases with possible early ectatic disease is essential for laser vision correction (not always performed by skilled cornea experts), what makes critical to support clinicians with automatic methods for detecting keratoconus. We have proven that the addition of new OCT based epithelial and stromal thickness indexes improves ANN detection capacity of keratoconus suspect eyes. For already stablished keratoconus our ANN detection capacity is excellent, but equivalent to previous evidence without incorporating such new OCT-based indexes.