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

The Evolution of Diagnostics for Keratoconus: From Ophthalmometry to Biomechanics

PURPOSE

To enumerate the various diagnostic modalities used for keratoconus and their evolution over the past century.

METHODS

A comprehensive literature search including articles on diagnosis on keratoconus were searched on PUBMED and summarized in this review.

RESULTS

Initially diagnosed in later stages of the disease process through clinical signs and retinoscopy, the initial introduction of corneal topography devices like Placido disc, photokeratoscopy, keratometry and computer-assisted videokeratography helped in the earlier detection of keratoconus. The evolution of corneal tomography, initially with slit scanning devices and later with Scheimpflug imaging, has vastly improved the accuracy and detection of clinical and sub-clinical disease. Analyzing the alteration in corneal biomechanics further contributed to the earlier detection of keratoconus even before the tomographic changes became evident. Anterior segment optical coherence tomography has proven to be a helpful adjuvant in diagnosing keratoconus, especially with epithelial thickness mapping. Confocal microscopy has helped us understand the alterations at a cellular level in keratoconic corneas.

CONCLUSION

Thus, the collective contribution of the various investigative modalities have greatly enhanced earlier and accurate detection of keratoconus, thus reducing the disease morbidity.

Use of machine learning to achieve keratoconus detection skills of a corneal expert

PURPOSE

To construct an automatic machine-learning derived algorithm discriminating between normal corneas and suspect irregular or keratoconic corneas.

METHODS

A total of 8526 corneal tomography images of 4904 eyes obtained between November 2010 and July 2017 using a combined Scheimpflug/Placido tomographer were retrospectively evaluated. Each image was evaluated for acquisition quality and was labeled as normal, suspect irregular or keratoconic by a cornea specialist. Two algorithms were built. The first was based on 94 instrument-derived output parameters, and the second integrated keratoconus prediction indices of the device with the 94 instrument-derived output parameters. Both models were compared with the tomographer's keratoconus detection algorithms. Out of the 8526 images evaluated, 7104 images of 3787 eyes had sufficient acquisition quality. Of those, 5904 examinations were randomly chosen for construction of the models using the random forest algorithm. The models were then validated using the remaining 1200 examinations.

RESULTS

Both RF algorithms had a larger AUC compared with any of the tomographer's KC detection algorithms (p < 10^-9). The first constructed model had 90.2% accuracy, sensitivity of 94.2%, and specificity of 89.6% (Youden 0.838). Calculated AUC was 0.964. The second model had 91.5% accuracy, sensitivity of 94.7%, and specificity of 89.8% (Youden 0.846). Calculated AUC was 0.969.

CONCLUSION

Using the RF machine-learning algorithm, accuracy of discrimination between normal, suspect irregular and keratoconic corneas approximates that of an experienced corneal expert. Applying machine learning to corneal tomography can facilitate keratoconus screening in large populations as well as off-site screening of refractive surgery candidates.

Screening Candidates for Refractive Surgery With Corneal Tomographic-Based Deep Learning

Importance

Evaluating corneal morphologic characteristics with corneal tomographic scans before refractive surgery is necessary to exclude patients with at-risk corneas and keratoconus. In previous studies, researchers performed screening with machine learning methods based on specific corneal parameters. To date, a deep learning algorithm has not been used in combination with corneal tomographic scans.

Objective

To examine the use of a deep learning model in the screening of candidates for refractive surgery.

Design, Setting, and Participants

A diagnostic, cross-sectional study was conducted at the Zhongshan Ophthalmic Center, Guangzhou, China, with examination dates extending from July 18, 2016, to March 29, 2019. The investigation was performed from July 2, 2018, to June 28, 2019. Participants included 1385 patients; 6465 corneal tomographic images were used to generate the artificial intelligence (AI) model. The Pentacam HR system was used for data collection.

Interventions

The deidentified images were analyzed by ophthalmologists and the AI model.

Main Outcomes and Measures

The performance of the AI classification system.

Results

A classification system centered on the AI model Pentacam InceptionResNetV2 Screening System (PIRSS) was developed for screening potential candidates for refractive surgery. The model achieved an overall detection accuracy of 94.7% (95% CI, 93.3%-95.8%) on the validation data set. Moreover, on the independent test data set, the PIRSS model achieved an overall detection accuracy of 95% (95% CI, 88.8%-97.8%), which was comparable with that of senior ophthalmologists who are refractive surgeons (92.8%; 95% CI, 91.2%-94.4%) (P = .72). In distinguishing corneas with contraindications for refractive surgery, the PIRSS model performed better than the classifiers (95% vs 81%; P < .001) in the Pentacam HR system on an Asian patient database.

Conclusions and Relevance

PIRSS appears to be useful in classifying images to provide corneal information and preliminarily identify at-risk corneas. PIRSS may provide guidance to refractive surgeons in screening candidates for refractive surgery as well as for generalized clinical application for Asian patients, but its use needs to be confirmed in other populations.

Characterization of cone size and centre in keratoconic corneas

A novel method to locate the centre of keratoconus (KC) and the transition zone between the pathological area and the rest of the corneal tissue is proposed in this study. A spherical coordinate system was used to generate a spherical height map measured relative to the centre of the optimal sphere fit, and normal to the surface. The cone centre was defined as the point with the maximum height. Second derivatives of spherical height were then used to estimate the area of pathology in an iterative process. There was mirror symmetry between cone centre locations in both eyes. The mean distance between cone centre and corneal apex was 1.45 ± 0.25 mm (0.07-2.00), the mean cone height normal to the surface was 37 ± 23 µm (2-129) and 75 ± 45 µm (5-243) in the anterior and posterior surfaces, respectively. There was a significant negative correlation between the cone height and the radius of the sphere of optimal fit (p < 0.05 for both anterior and posterior surfaces). On average, posterior cone height was larger than the corresponding anterior cone height by 37 ± 24 µm (0-158). The novel method proposed can be used to estimate the cone centre and area, and explore the changes in anterior and posterior corneal surfaces that take place with KC progression. It can help improve understanding of keratoconic corneal morphology and assist in developing customized treatments.

A Review of Machine Learning Techniques for Keratoconus Detection and Refractive Surgery Screening

Various machine learning techniques have been developed for keratoconus detection and refractive surgery screening. These techniques utilize inputs from a range of corneal imaging devices and are built with automated decision trees, support vector machines, and various types of neural networks. In general, these techniques demonstrate very good differentiation of normal and keratoconic eyes, as well as good differentiation of normal and form fruste keratoconus. However, it is difficult to directly compare these studies, as keratoconus represents a wide spectrum of disease. More importantly, no public dataset exists for research purposes. Despite these challenges, machine learning in keratoconus detection and refractive surgery screening is a burgeoning field of study, with significant potential for continued advancement as imaging devices and techniques become more sophisticated.

CLMI: the cone location and magnitude index

PURPOSE

To develop an index for the detection of keratoconic patterns in corneal topography maps from multiple devices.

METHODS

For development, an existing Keratron (EyeQuip) topographic dataset, consisting of 78 scans from the right eyes of 78 healthy subjects and 25 scans from the right eyes of 25 subjects with clinically diagnosed keratoconus, was retrospectively analyzed. The Cone Location and Magnitude Index (CLMI) was calculated on the available axial and tangential curvature data. Stepwise logistic regression analysis was performed to determine the best predictor(s) for the detection of keratoconus. A sensitivity and specificity analysis was performed by using the best predictor of keratoconus. Percent probability of keratoconus was defined as the optimal probability threshold for the detection of disease. For validation, CLMI was calculated retrospectively on a second distinct dataset, acquired on a different topographer, the TMS-1. The validation dataset consisted of 2 scans from 24 eyes of 12 healthy subjects with no ocular history and 4 scans from 21 eyes of 14 subjects with clinically diagnosed keratoconus. Probability of keratoconus was calculated for the validation set from the equation determined from the development dataset.

RESULTS

The strongest significant sole predictor in the stepwise logistic regression was aCLMI, which is CLMI calculated from axial data. Sensitivity and specificity for aCLMI on the development dataset were 92% and 100%, respectively. A complete separation of normals and keratoconics with 100% specificity and 100% sensitivity was achieved by using the validation set.

CONCLUSIONS

CLMI provides a robust index that can detect the presence or absence of a keratoconic pattern in corneal topography maps from 2 devices.

Spatial Modeling and Classification of Corneal Shape

One of the most promising applications of data mining is in biomedical data used in patient diagnosis. Any method of data analysis intended to support the clinical decision-making process should meet several criteria: it should capture clinically relevant features, be computationally feasible, and provide easily interpretable results. In an initial study, we examined the feasibility of using Zernike polynomials to represent biomedical instrument data in conjunction with a decision tree classifier to distinguish between the diseased and non-diseased eyes. Here, we provide a comprehensive follow-up to that work, examining a second representation, pseudo-Zernike polynomials, to determine whether they provide any increase in classification accuracy. We compare the fidelity of both methods using residual root-mean-square (rms) error and evaluate accuracy using several classifiers: neural networks, C4.5 decision trees, Voting Feature Intervals, and Naïve Bayes. We also examine the effect of several meta-learning strategies: boosting, bagging, and Random Forests (RFs). We present results comparing accuracy as it relates to dataset and transformation resolution over a larger, more challenging, multi-class dataset. They show that classification accuracy is similar for both data transformations, but differs by classifier. We find that the Zernike polynomials provide better feature representation than the pseudo-Zernikes and that the decision trees yield the best balance of classification accuracy and interpretability.

A New Method for Grading the Severity of Keratoconus

PURPOSE

To define a new method for grading severity of keratoconus, the Keratoconus Severity Score (KSS).

METHODS

A rationale for grading keratoconus severity was developed using common clinical markers plus 2 corneal topographic indices, creating a 0 to 5 severity score. An initial test set of 1012 eyes, including normal eyes, eyes with abnormal corneal and topographic findings but not keratoconus, and eyes with keratoconus having a wide range of severity, was used to determine cutpoints for the KSS. Validation set 1, comprising data from 128 eyes, was assigned a KSS and compared with a clinician's ranking of severity termed the "gold standard" to determine if the scale fairly represented how a clinician would grade disease severity. kappa statistics, sensitivity, and specificity were calculated. A program was developed to automate the determination of the score. This was tested against a manual assignment of KSS in 2121 (validation set 2) eyes from the Collaborative Longitudinal Evaluation of Keratoconus (CLEK) Study, as well as normal eyes and abnormal eyes without keratoconus. Ten percent of eyes underwent repeat manual assignment of KSS to determine the variability of manual assignment of a score.

RESULTS

From initial assessments, the KSS used 2 corneal topography indices: average corneal power and root mean square (RMS) error for higher-order Zernike terms derived from the first corneal surface wavefront. Clinical signs including Vogt striae, Fleischer rings, and corneal scarring were also included. Last, a manual interpretation of the map pattern was included. Validation set 1 yielded a kappa statistic of 0.904, with sensitivities ranging from 0.64 to 1.00 and specificities ranging from 0.93 to 0.98. The sensitivity and specificity for determining nonkeratoconus from keratoconus were both 1.00. Validation set 2 showed kappa statistics of 0.94 and 0.95 for right and left eyes, respectively. Test-retest analysis yielded kappa statistics of 0.84 and 0.83 for right and left eyes, respectively.

CONCLUSION

A simple and reliable grading system for keratoconus was developed that can be largely automated. Such a grading scheme could be useful in genetic studies for a complex trait such as keratoconus requiring a quantitative measure of disease presence and severity.

Preliminary results of neural networks and zernike polynomials for classification of videokeratography maps

PURPOSE

Our main goal in this work was to develop an artificial neural network (NN) that could classify specific types of corneal shapes using Zernike coefficients as input. Other authors have implemented successful NN systems in the past and have demonstrated their efficiency using different parameters. Our claim is that, given the increasing popularity of Zernike polynomials among the eye care community, this may be an interesting choice to add complementing value and precision to existing methods. By using a simple and well-documented corneal surface representation scheme, which relies on corneal elevation information, one can generate simple NN input parameters that are independent of curvature definition and that are also efficient.

METHODS

We have used the Matlab Neural Network Toolbox (MathWorks, Natick, MA) to implement a three-layer feed-forward NN with 15 inputs and 5 outputs. A database from an EyeSys System 2000 (EyeSys Vision, Houston, TX) videokeratograph installed at the Escola Paulista de Medicina-Sao Paulo was used. This database contained an unknown number of corneal types. From this database, two specialists selected 80 corneas that could be clearly classified into five distinct categories: (1) normal, (2) with-the-rule astigmatism, (3) against-the-rule astigmatism, (4) keratoconus, and (5) post-laser-assisted in situ keratomileusis. The corneal height (SAG) information of the 80 data files was fit with the first 15 Vision Science and it Applications (VSIA) standard Zernike coefficients, which were individually used to feed the 15 neurons of the input layer. The five output neurons were associated with the five typical corneal shapes. A group of 40 cases was randomly selected from the larger group of 80 corneas and used as the training set.

RESULTS

The NN responses were statistically analyzed in terms of sensitivity [true positive/(true positive + false negative)], specificity [true negative/(true negative + false positive)], and precision [(true positive + true negative)/total number of cases]. The mean values for these parameters were, respectively, 78.75, 97.81, and 94%.

CONCLUSION

Although we have used a relatively small training and testing set, results presented here should be considered promising. They are certainly an indication of the potential of Zernike polynomials as reliable parameters, at least in the cases presented here, as input data for artificial intelligence automation of the diagnosis process of videokeratography examinations. This technique should facilitate the implementation and add value to the classification methods already available. We also discuss briefly certain special properties of Zernike polynomials that are what we think make them suitable as NN inputs for this type of application.

Wavelet analysis for corneal topographic surface characterization

PURPOSE

To demonstrate a mathematical method for multiscalar decomposition of discrete corneal topography height data into a space-scale space using wavelet analysis techniques, and to demonstrate the clinical applicability of these computations in the postkeratoplasty cornea.

METHODS

Fifty patients with either Fuchs' dystrophy (n = 20) or keratoconus (n = 30) were seen preoperatively, at 3 months, at 1 year (before suture removal) and again at 19 +/- 3 months (after suture removal) following nonmechanical trephination with an excimer laser for penetrating keratoplasty. Patients were assessed using corneal topography analysis (TMS-1), subjective refraction, and best-corrected visual acuity (VA) at each interval. Two-dimensional biorthogonal wavelets with the order 6.8 at the scales j = 1-4 revealed the following parameters: root-mean square (RMSDEV) and mean absolute (MEANDEV) deviation and maximum absolute height of the peaks or pitches (MAXPEAK) relative to the reference surface specified with the approximation component of scale j = 4. RMSDEV was correlated with the VA at various follow-up intervals. The multiscalar basis components: roughness, waviness and form were separated and recovered from the wavelet soft thresholding techniques. Peaks and pits within the three-dimensional corneal surface topography were detected and localized using the wavelet hard thresholding techniques.

RESULTS

In patients with keratoconus, the RMSDEV and the MEANDEV increased from 4.31 +/- 1.25/5.98 +/- 1.88 microm preoperatively to the 3 months follow-up (4.98 +/- 1.41/6.92 +/- 2.16 microm) and thereafter decreased continuously to the end of the follow-up (1.87 +/- 0.63/2.63 +/- 1.07 microm), whereas in Fuchs' dystrophy the respective values started at a higher preoperative level (6.36 +/- 1.24/7.20 +/- 2.64 microm) and decreased continuously over time (2.73 +/- 1.10/3.71 +/- 1.05 microm after suture removal). In the keratoconus group, the MAXPEAK was increased at the 3 month postoperative exam (8.78 +/- 2.29 microm) when compared to the preoperative value (6.55 +/- 2.56 microm); however, it decreased again and returned to the preoperative level after one year (6.34 +/- 2.12 microm after suture removal). In Fuchs' dystrophy, the MAXPEAK was unchanged preoperatively (8.26 +/- 2.83 microm) to the 3 months follow-up, but decreased continuously to the end of the follow-up period (4.57 +/- 1.36 microm). The RMSDEV was significantly lower in keratoconus than in Fuchs' dystrophy preoperatively (P = 0.01) and after suture removal (P = 0.005) and correlated inversely with VA preoperatively (R = -0.53, P = 0.04/R = -0.69, P = 0.02), at the 1 year exam (R = -0.61, P = 0.02/R = -0.52, P = 0.05) and after suture removal (R = -0.73, P = 0.01/R = -0.66, P = 0.025) in keratoconus/Fuchs' dystrophy.

CONCLUSIONS

The use of wavelet analysis can provide significant clinical information by separating the raw data into the parameters: "roughness", "waviness", "form" and various multiscalar peaks and pits. The RMSDEV, a quantitative measure for corneal irregularity, can be used as a new approach for the prediction of potential visual acuity after penetrating keratoplasty. The decomposition of the surface elevation into fundamental components is crucial for a subsequent mathematically based extraction of clinical parameters or for topography-based flying-spot ablation of irregular corneal astigmatism.