Artificial Intelligence for Pediatric Ophthalmology

Julia E. Reid, Eric EatonarXiv 2019 · OpenReview

Abstract

Despite the impressive results of recent artificial intelligence (AI) applications to general ophthalmology, comparatively less progress has been made toward solving problems in pediatric ophthalmology. This article discusses the unique needs of pediatric patients, surveys recent AI applications, and discusses future directions. Significant advances involve automated detection of retinopathy of prematurity (ROP), yielding results that rival experts. Machine learning (ML) has also been applied to pediatric cataracts, strabismus, refractive error, and visual development studies. AI could significantly benefit clinical care by optimizing disease detection, broadening access, and improving efficiency, though methods must match physician performance in clinical trials before deployment.

The Unique Challenges of Pediatric Eye Care

The "AI boom" in medicine has largely focused on adult populations. In ophthalmology, deep learning has achieved human-level performance in detecting diabetic retinopathy and glaucoma. However, pediatric ophthalmology faces distinct hurdles. Children are not simply "small adults"; they suffer from different conditions like Retinopathy of Prematurity (ROP) and congenital cataracts.

Furthermore, pediatric eyes undergo rapid physiological changes during development, and children often have shorter attention spans, making high-quality imaging difficult. There is also a critical global shortage of pediatric ophthalmologists. AI could bridge this gap by providing automated screening and decision support in regions where specialists are unavailable.

Methodologies and Data Scarcity

Convolutional Neural Networks (CNNs) are the dominant architecture for image-based diagnosis in this field. Models like Inception and ResNet are frequently used, often via transfer learning to compensate for small pediatric datasets. Some systems, like i-ROP-DL, utilize hybrid pipelines where a U-Net performs automated vessel segmentation before a classification CNN analyzes the data.

Figure 1: Examples of pediatric fundus images and their corresponding automated vessel segmentations, a key step in many AI diagnostic pipelines.

Data scarcity remains a primary bottleneck. Unlike adult ophthalmology, which benefits from large public datasets, pediatric researchers often work with "closed-access" sets. This makes it difficult to reproduce results or compare different algorithms directly. To combat this, researchers are exploring Generative Adversarial Networks (GANs) to synthesize realistic fundus images, providing more training material without compromising patient privacy.

Breakthroughs in Disease Detection

The most significant progress has been made in screening for Retinopathy of Prematurity (ROP). The DeepROP system achieved over 96% sensitivity for detecting the disease, while i-ROP-DL reached an AUROC of 0.98 for "plus disease"—a critical marker where even human experts often disagree.

Figure 5: Summary of various AI approaches and their performance metrics for detecting ROP plus and pre-plus disease.

Beyond ROP, AI is showing promise in other areas:

  • Pediatric Cataracts: The CC-Cruiser platform achieved 87.4% accuracy in multi-center trials.
  • Refractive Error: Random forest models have predicted the onset of high myopia in children up to eight years in advance.
  • Strabismus: CNNs using simple facial photos have achieved over 93% accuracy, highlighting potential for telemedical screening.

Future Directions and Limitations

Despite these successes, several challenges remain. Most deep learning models are "black boxes," providing little explanation for their diagnoses—a significant hurdle for clinical trust. Additionally, most current AI looks at a single "snapshot" rather than longitudinal data that tracks how a child's eye changes over time.

Future progress will require more open-access datasets and software implementations to improve reproducibility. There is also a dire need for prospective clinical trials to validate these tools in real-world workflows. If these hurdles are cleared, AI could revolutionize pediatric eye care by making expert-level screening accessible to children worldwide.

Summary updated

Abstract

PURPOSE OF REVIEW: Despite the impressive results of recent artificial intelligence (AI) applications to general ophthalmology, comparatively less progress has been made toward solving problems in pediatric ophthalmology using similar techniques. This article discusses the unique needs of pediatric ophthalmology patients and how AI techniques can address these challenges, surveys recent applications of AI to pediatric ophthalmology, and discusses future directions in the field. RECENT FINDINGS: The most significant advances involve the automated detection of retinopathy of prematurity (ROP), yielding results that rival experts. Machine learning (ML) has also been successfully applied to the classification of pediatric cataracts, prediction of post-operative complications following cataract surgery, detection of strabismus and refractive error, prediction of future high myopia, and diagnosis of reading disability via eye tracking. In addition, ML techniques have been used for the study of visual development, vessel segmentation in pediatric fundus images, and ophthalmic image synthesis. SUMMARY: AI applications could significantly benefit clinical care for pediatric ophthalmology patients by optimizing disease detection and grading, broadening access to care, furthering scientific discovery, and improving clinical efficiency. These methods need to match or surpass physician performance in clinical trials before deployment with patients. Due to widespread use of closed-access data sets and software implementations, it is difficult to directly compare the performance of these approaches, and reproducibility is poor. Open-access data sets and software implementations could alleviate these issues, and encourage further AI applications to pediatric ophthalmology. KEYWORDS: pediatric ophthalmology, machine learning, artificial intelligence, deep learning
BibTeX
@misc{reid2019artificial,
  author = {Julia E. Reid and Eric Eaton},
  title = {{Artificial Intelligence for Pediatric Ophthalmology}},
  year = {2019},
  eprint = {1904.08796},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1904.08796},
}
Figure