Artificial Intelligence for Pediatric Ophthalmology

Julia E. Reid; Eric Eaton

Curr Opin Ophthalmol. 2019;30(5):337-346.

Conclusion

There is a large potential for current and future artificial intelligence applications to pediatric ophthalmology, and there are some diseases, such as NLDO, congenital glaucoma, and congenital ptosis, without any published applications of artificial intelligence to our knowledge. Automated disease detection, the most common use case, could augment telemedical efforts to broaden access to care, improve efficiency, and result in earlier diagnoses. However, other less-utilized capabilities of this technology, including disease grading and outcome prediction, have the potential to enhance clinical care. All artificial intelligence methods deployed in clinical care must ultimately match or surpass physician performance while meeting the unique requirements of both clinicians and pediatric patients, suggesting the need to augment evaluations on experimental data sets with clinical trials.

1 2 3 4 5 6

Next Section

Acknowledgements
We would like to thank Jing Jin, MD, José Marcio Luna, PhD, and Jorge Mendez for their helpful feedback on this article.

Financial support and sponsorship
E.E.'s work was partially supported by the Lifelong Learning Machines program from DARPA/MTO under grant number FA8750-18-2-0117. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of DARPA or the U.S. Government.

Table 1. Summary of machine learning–based techniques for pediatric ophthalmic disease detection and diagnosis
Approach (Approx. devel. year)	Predicted category	Sensitivity (%)	Specificity (%)	AUROC	Accuracy (%)	Method summary
Retinopathy of Prematurity (ROP)
DeepROP [17] (2018)	Experimental data set					Cloud-based platform. Set of fundus images!two CNNs (modified inception-BN nets pretrained on ImageNet): one predicts presence and the other severity
	Presence of ROP	96.64	99.33	0.995	97.99
	Severe (versus mild) ROP	88.46	92.31	0.951	90.38
	Clinical test
	Presence of ROP	84.91	96.90	–	95.55
	Severe (versus mild) ROP	93.33	73.63	–	76.42
i-ROP-DL [18] (2018)	Clinically significant ROP	–	–	0.914	–	Applies a linear formula to the probabilities output by i-ROP-DL (see below) to yield a severity score on a 1–9 scale
	Type 1 ROP	94	79	0.960	–
	Type 2 ROP	–	–	0.867	–
	Pre-plus disease	–	–	0.910	–
MiGraph [19] (2016)	Presence of ROP	99.4	95	0.98	97.5	SIFT features from image patches→multiple instance learning graph-kernel SVM
VesselMap [20] (2007)	Severe ROP					Semiautomated tool that uses classic image analysis to measure vessel diameter
	From mean arteriole diameter	–	–	0.93	–
	From mean venule diameter	–	–	0.87	–
ROP: Plus or pre-plus disease
i-ROP-DL [21] (2018)	Plus disease [18]	–	–	0.989	–	CNN-output (U-net) vessel segmentations→CNN (Inception V1 pretrained on ImageNet) to classify as normal/pre-plus/plus
	Pre-plus disease [18]	–	–	0.910	–
	Plus disease [21]	93	94	0.98	91.0
	Pre-plus or worse disease [21]	100	94	0.94	–
CNN+Bayes [16] (2016)	Plus disease (per image)	82.5	98.3	–	91.8	CNN (Inception V1 pretrained on ImageNet) adapted to output the Bayesian posterior
CNN+Bayes [16] (2016)	Plus disease (per examination)	95.4	94.7	–	93.6
i-ROP [22] (2015)	Plus disease	93	–	–	95	SVM with a kernel derived from a GMM of tortuosity and dilation features from manually segmented images
i-ROP [22] (2015)	Pre-plus or worse disease	97	–	–	–
Naïve Bayes [23] (2015)	Plus/pre-plus/none (SVM-RFE)	–	–	–	79.41	Naïve Bayes with SVM-RFE or ReliefF vessel feature selection
Naïve Bayes [23] (2015)	Plus disease (ReliefF)	–	–	–	88.24
CAIAR [24] (2008)	Plus (from venule width)	–	–	0.909	–	Generative vessel model fit to a multiscale representation of the retinal image
CAIAR [24] (2008)	Plus (from arteriole tortuosity)	–	–	0.920	–
ROPtool [26] (2007)	Plus tortuosity (eye)	95	78	–	87.50	User-guided tool that traces centerlines of retinal vessels to measure tortuosity
	Plus tortuosity (quadrant)	85	77	0.885	80.63
	Pre-plus tortuosity (quadrant)	89	82	0.875	–
RISA [27] (2005)	Plus disease (from arteriole and venule curvature and tortuosity, venule diameter)	93.8	93.8	0.967	–	Logistic regression on geometric features computed for each segment of the vascular tree
IVAN [24] (2002)	Plus (from venule width)	–	–	0.909	–	Measures vessel width via classic image analysis
Pediatric cataracts
Postoperative complication prediction [28] (2019)	CLR and/or high IOP (random forest)	62.5	76.9	0.722	70.0	Demographic and cataract severity evaluation data→classbalancing using SMOTE!random forest and naïve Bayes classifiers
	CLR and/or high IOP (naïve Bayes)	73.1	66.7	0.719	70.0
	Central lens regrowth (random forest)	66.7	72.2	0.743	72.0
	Central lens regrowth (naïve Bayes)	61.1	68.8	0.735	66.0
	High IOP (random forest	63.6	71.8	0.735	70.0
	High IOP (naïve Bayes)	54.5	69.2	0.719	66.0
CS-ResCNN [29] (2017)	Severe posterior capsular opacification	89.66	93.19	0.9711	92.24	Slit-lamp images→automatically crop to lens→CNN (ResNet pretrained on ImageNet) with cost-sensitive loss
CC-Cruiser [30] (2016)	Multicenter trial					Cloud-based platform. Slit-lamp images→automatically crop to lens→three CNNs (AlexNets) to predict: cataract presence, severity (area, density, location), and treatment (surgery or follow-up)
	Cataract presence [31]	89.7	86.4	–	87.4
	Opacity area grading [31]	91.3	88.9	–	90.6
	Density grading [31]	85.3	67.9	–	80.2
	Location grading [31]	84.2	50.0	–	77.1
	Treatment [31]	86.7	44.4	–	70.8
	Experimental data set
	Cataract presence [32]	96.83	97.28	0.9686	97.07
	Area grading [32]	90.75	86.63	0.9892	89.02
	Density grading [32]	93.94	91.05	0.9743	92.68
	Location grading [32]	93.08	82.70	0.9591	89.28
Strabismus
RF-CNN [33] (2018)	Strabismus presence	93.30	96.17	0.9865	93.89	Two-stage CNN: eye regions segmented from face images via R-FCN→11-layer CNN
SVM+VGG-S [34] (2017)	Strabismus presence	94.1	96.0	–	95.2	Eye-tracking gaze maps→CNN (VGG-S pretrained on ImageNet) features→SVM
Pediatric vision Screener [35] (2017)	Central versus paracentral fixation					Signals from retinal birefringence scanning→two-layer feed-forward neural net
	Experimental evaluation	100.0	100.0	–	–
	Clinical evaluation	98.51	100.0	–	–
Vision screening
AVVDA [36] (2008)	Strabismus and/or refractive error	–	–	–	76.9	Features from Brückner red reflex imaging and eccentric fixation video→C4.5 decision tree
	Strabismus	82	–	–	–
	High refractive error	90	–	–	–
Reading disability
SVM-RFE [37] (2016)	High risk for reading disability, ages 8–9	95.5	95.7	–	95.6	SVM with feature selection trained on eye-tracking data
Polynomial SVM [38] (2015)	Reading disability in adults, children ages 11+	–	–	–	80.18	SVM trained on eye-tracking and demographic features
(Approx. devel. year)	Predicted category	AUROC (at 3 years)	AUROC (at 5 years)		AUROC (at 8 years)	Method summary
Refractive error
Random forest [39] (2018)	Internal evaluation					Age, spherical equivalent, and progression rate of spherical equivalent between two visits was used by a random forest for prediction
	High myopia onset	0.903–0.986	0.875–0.901		0.852–0.888
	Clinical test
	High myopia onset	0.874–0.976	0.847–0.921		0.802–0.886
	High myopia at age 18	0.940–0.985	0.856–0.901		0.801–0.837

AVVDA, Automated Video Vision Development Assessment; AUROC, Area under the receiver operating characteristic curve; CAIAR, computer-assisted image analysis of the retina; CNN, convolutional neural network; GMM, Gaussian mixture model; IOP, intraocular pressure; RFE, recursive feature elimination; SVM, support vector machine.

Table 2. Pediatric ROP data sets used in deep learning
Approach	Data set	Patients	Images	Labels
DeepROP [17]	Chengdu	1273	20795	Normal, mild ROP, severe ROP
i-ROP-DL [21]	i-ROP	898	5511	Normal, plus, pre-plus
CNN+Bayes [16]	Canada	35	1459	Normal, plus
	London	–	106	Normal, plus