Purchase this article with an account.
M.K. Smolek, S.D. Klyce, M.D. Karon; Device–Independent Corneal Topography Classification and Keratoconus Grading by Neural Networks . Invest. Ophthalmol. Vis. Sci. 2004;45(13):2867.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose: Previously, automated corneal topography classification within a universal scheme was impossible due to spatial resolution and mire differences among various topographers. A newly developed 2–D fast Fourier transform (FFT) now allows feature index analysis for placido topographers at a standard spatial resolution. Artificial neural networks (NN) were made to automatically classify device–independent maps and grade the severity of keratoconus. Methods: A total of 1825 OPD–Scan topography maps were acquired in 8 categories: normal, astigmatic, keratoconus suspect, keratoconus (KC), pellucid marginal degeneration, penetrating keratoplasty, and hyperopic and myopic refractive surgery. After review, 224 maps were chosen for NN training based on image quality and features deemed typical for each category. Another 224 maps were chosen at random from each category for testing. Each map was Fourier–filtered to a common resolution and 18 corneal indexes extracted for use as NN inputs (SimK1, SimK2, MinK, CYL, SRI, SRC, OSI, DSI, CSI, SAI, CEI, IAI, ACP, AA, SDP, CVP, KPI, & KCI). These corneal indices were previously shown to be sensitive to various topographical features. Not all NNs required all 18 inputs; some required as few as 11 inputs. A separate NN was trained and tested for each category. The NN outputs were then passed through an expert system threshold filter to obtain the final classification score. Maps classified as KC were also used to train and test a KC severity NN that produced a numerical index to distinguish mild, moderate, and severe forms. Results: In the independent test set, the classification network scored 91% of maps with a single correct response. An additional 3% of maps elicited 2 responses, with the second highest response being correct. 6% of maps were classified as unknown and these were mainly keratoplasty cases. KC severity output was 100% correct with the majority of test set maps displaying <5% error from their intended severity grade. Subsequently, the NNs have been tested with maps from various topographers (TMS, Humphrey, Magellan, etc) and have shown excellent performance for both classification and KC grading. Conclusions: Fourier filtering makes it possible to classify maps from any corneal topographer with NNs, no matter what camera resolution or mire pattern was used. Likewise, keratoconus severity can be graded correctly and identically on a map from any system. Clinicians will benefit from automated interpretation that can be transferred among various systems and that allows maps from different systems to be directly compared by corneal statistics or neural networks.
This PDF is available to Subscribers Only