Screening of prior refractive surgery by a wavelet-based neural network

Purpose: To demonstrate an objective method of screening for previous refractive surgery using corneal topography. Setting: Corneal topography research laboratories, LSU Eye Center, New Orleans, Louisiana, USA, Methods: Videokeratography (TMS-1, Tomey) examinations from the LSU Eye Center were randomly divided into training and test sets that each included 32 normal corneas and 106 corneas with previous radial keratotomy or photorefractive keratectomy from 1 month up to 10 years after surgery, A set of 1024 axial curvature values were extracted from mires 1 to 25 from each cornea to form a 1-dimensional waveform. Multiresolution wavelet decomposition was performed on this waveform using the s8 Symmlet wavelet. A portion of the resulting wavelet coefficients was input into a backpropagation neural network that was trained to 5% error. After training, the independent test set was passed though the neural net and scored. Results: The trained network correctly recognized 32 of 32 normal corneas and 105 of 106 refractive surgery corneas for a 99.3% accuracy, 99.1% sensitivity, and 100% specificity for previous myopic refractive surgery detection. Conclusions: The 1-dimensional wavelet-based neural network approach was an effective and accurate method of distinguishing eyes that had had myopic refractive surgery from normal eyes. The single error was a result of having too few examples of grossly decentered procedures in the training set. (C) 2001 ASCRS and ESCRS.