Table 3. Artificial Intelligence Mathematical Models Used on Videokeratographic and Corneal Topographic or Tomographic Imaging.
| Source | Algorithm | Instrument | Input | Output | Data set, No. | Performance | |
|---|---|---|---|---|---|---|---|
| Training | Test/validation | ||||||
| Maeda et al,101994 | KPI (linear discriminant analysis) | TMS-1 videokeratoscope | 8 Topographic indices | KC, non-KC | 100 | 100 | Accuracy, 96%; sensitivity, 89%; specificity, 99% (cutoff value, 0.23) |
| Maeda et al,12 1995 | Neural network | TMS-1 videokeratoscope | 11 Topographic indices | Normal, WA, KC1, KC2, KC3, post-PRK, post-KP | 108 | 75 | Total accuracy, 80%; KC accuracy, 92%-97% |
| Smolek and Klyce,13 1997 | KSI (neural network plus binary decision tree) | TMS-1 videokeratoscope | 10 Topographic indices | Other, KCS, KC1, KC2, KC3 | 150 | 150 | Total accuracy, 100%; KC accuracy, 100% |
| Smolek and Klyce,14 2001 | Neural network | TMS-1 videokeratoscope | Wavelet data | Normal, post-RS | 138 | 138 | Total accuracy, 99.3% |
| Accardo and Pensiero,15 2002 | Neural network | EyeSys 2000 | 9 Topographic indices | Normal, KC, others | 95 | 103 | Sensitivity, 94.1%; specificity, 97.6% |
| Twa et al,19 2005 | Decision tree | Keratron corneal topographer | Seventh-order Zernike polynomial | Normal, KC | 244 | Cross-validation | Total accuracy, 92% |
| Klyce et al,16 2005 | Corneal navigator (neural network) | OPD scanning system | 19 topographic indices | Normal, WA, KCS, KC, PMD, post-KP, HRS, post-RS, others | NA | NA | Unavailable |
| Vieira de Carvalho and Barbosa,39 2008 | Neural network plus discriminant analysis | EyeSys 2000 | Zernike coefficients | KC, WA, AA, post-PRK | 40 | 40 | Total accuracy: 94% by NN, 85% by DA |
| Souza et al,21 2010 | Neural network, support vector machine, multilayer perception | Orbscan system | 11 Topographic indices | KC, astigmatism, post-PRK | 318 | Cross-validation | AUC of KC: 0.99 by NN, 0.99 by SVM, 0.99 by MLP |
| Saad and Gatinel,11 2012 | SCORE Analyzer (linear discriminant analysis) | Orbscan system | >10 Placido and tomographic indices | Normal, at risk for LASIK | NA | 183 | Sensitivity, 92%; specificity, 96% |
| Arbelaez et al,222012 | Support vector machine | Sirius system | Indices based on curvature, thickness, and height data of both the anterior and posterior corneal surface and pachymetry | Normal, subclinical KC, KC, abnormal | 800 | 2702 | Total accuracy, 95.9%; normal accuracy, 98.1%; subclinical accuracy, 97.3%; KC accuracy, 98.2%; abnormal accuracy, 98.3% |
| Smadja et al,20 2013 | Decision tree | Galilei system | 55 Parameters derived from anterior and posterior corneal measurements | Normal, FFKC, KC | 372 | Cross-validation | Sensitivity of KC and FFKC: KC vs normal cornea, 100%; FFKC vs normal cornea, 93.6% |
| Kovács et al17 2016 | Neural network | Pentacam HR system | Unilateral and bilateral indices | Normal, FFKC, KC | 135 | Cross-validation | Sensitivity of KC and FFKC: KC vs normal cornea, 100%; FFKC vs normal cornea, 90% |
| Ruiz Hidalgo et al,23 2016 | KA (support vector machine) | Pentacam HR system | 22 Parameters of the corneal curvature, eccentricity, anterior chamber, corneal volume, and pachymetry | Normal, KC, FFKC, AST, post-RS | 860 | Cross-validation | Total accuracy, 88.8%; sensitivity of KC and FFKC: KC vs normal cornea, 98.9%; FFKC vs normal cornea, 93.1% |
| Lopes et al,25 2018 | PRFI (random forest) | Pentacam HR system | >20 Keratometric values, topometric indices, and tomographic indices | Stable LASIK, PLE, KC, FFKC, normal | 3233 | 486 From stable, VAE-NT, and VAE-E | Sensitivity of VAE-NT and VAE-E: VAE-NT vs normal cornea, 85.2%; VAE-E vs normal cornea, 100% |
| Our model, 2019 | PIRSS | Pentacam HR system | Images of the corneal axial curvature, the front elevation, the back elevation, and the corneal thickness | Normal, suspect, early KC, KC, postoperative | 5130 | 1335 | Total accuracy, 95%; KC accuracy, 98% |
Abbreviations: AA, against-the-rule astigmatism; AST, astigmatism; AUC, area under the receiver operator characteristic curve; DA, discriminant analysis; FFKC, forme fruste keratoconus; HRS, hyperopic refractive surgery; KA, KC assistant; KC, keratoconus; KC1, mild keratoconus; KC2, moderate keratoconus; KC3, advanced keratoconus; KCS, suspected KC; KP, keratoplasty; KPI, KC predictive index; KSI, KC severity index; LASIK, laser in situ keratomileusis; MLP, multilayer perception; NA, not available; NN, neural network; PIRSS, Pentacam InceptionResNetV2 Screening System; PLE, post-LASIK ectasia; PMD, pellucid marginal degeneration; post-PRK, post–photorefractive keratectomy; post-KP, post–keratoplasty; post-RS, post–refractive surgery; PRFI, Pentacam random forest index; SVM, support vector machine; VAE-E, ectasia eye of very asymmetric cases; VAE-NT, normal topography eye of very asymmetric cases; WA, with-the-rule astigmatism.