Abstract
Purpose :
To establish and investigate reliability of a computerised machine learning algorithm to classify post orthokeratology contact lens wear corneal topography fitting patterns.
Methods :
Pre and post wear corneal topography maps measured using the Medmont E300 (Melbourne, Australia) from 935 overnight orthokeratology (OK) lens wearing eyes were classified by 3 experts in OK lens fitting using the Medmont Visual Studio software as either representing a Bulls Eye (BE), Smiley Face (SF), Central Island (CI) Frowny Face (FF), or Lateral Decentration (LD) pattern typically used to categorise OK wear outcomes. The same maps were exported in XML format and processed by subtracting post from pre wear tangential and axial curvature measurements to establish lens induced changes to corneal curvature. The extracted axial and tangential curvature values were in 50x50 cartesian format representing a 12mm XY grid of the measured corneal surfaces in uniform steps. Null values were then removed, and the remaining values downscaled by averaging from two 50x50 to two 20x20 cartesian grids. 75% of the maps from classified reduced dataset were randomly allocated to train a Multiclass Decision Forest (MDF) computerised machine learning (ML) algorithm with the remaining 25% used to evaluate the accuracy of the trained model.
Results :
Overall accuracy of the MDF ML classified maps in comparison to expert classification was 82.25% (p<0.001). 95.9% of BE, 73.9% of SF, 63.6% of CI, 40.0% of FF, and 58.3% of LD were correctly classified by the MDF ML algorithm. 21.7% of SF, 27.3% of CI, 60.0% of FF and 33.3% of LD were incorrectly classified as BE. BE maps were most likely to be correctly classified and FF least likely to be correctly classified.
Conclusions :
The results indicate that ML algorithms offer potential to assist practitioners in correctly classifying OK lens fit outcomes to more accurately direct subsequent improvements to lens fit. These algorithms would particularly benefit new OK lens fitters lacking experience evaluating post OK corneal topography changes. While highly accurate for classifying BE maps, and reasonably accurate for classifying SF, correct classification of CI and LD was only moderately successful. Further work is currently under way to improve classification accuracy by refining the feature extraction process and inclusion of more expert classified samples to increase the size of the dataset.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.