Deep Learning Offers Quantitative Means of Monitoring Disease Progression in Retinopathy of Prematurity


Gary Boas

Retinopathy of prematurity (ROP) is an eye disorder affecting roughly two-thirds of premature infants weighing less than 250 g at birth and one of the leading causes of childhood blindness worldwide. Historically, clinical diagnosis of ROP has been subjective, resulting in considerable variability in diagnosis and therefore in decisions about treatment. This is beginning to change. Last year, a team of researchers at the MGH Martinos Center and elsewhere described a deep learning algorithm that analyzes retinal images and provides a quantitative ROP “severity score” with comparable or better accuracy than clinical experts. Now, as reported in a JAMA Ophthalmology paper published online in July, the researchers have demonstrated use of the algorithm for objective monitoring of ROP progression. The study’s findings underscore the potential of deep learning to help identify infants at high risk for developing ROP.

Diagnosis of ROP has always been based on qualitative assessment. Since the 1980s, physicians have identified ROP by manually determining whether arterial tortuosity and venous dilation of the posterior retinal vessels in a diagnostic image are greater than or equal to those shown in a standard reference image. In 2005, an international panel introduced a three-tier grading classification system for plus disease (normal, pre-plus disease and plus disease), adding an intermediate level of severity that can aid with prognosis. Here, though, diagnosis of the disorder was still based on a subjective assessment.

Deep learning studies published over the past several years have begun to address the subjectivity problem. The 2018 report by the Martinos group and colleagues built on earlier work showing promising results for two-level diagnosis of plus disease in ROP using algorithms trained on a large dataset of ROP patients. In this study, they demonstrated use of a deep learning approach for three-level diagnosis – adding an intermediate level of severity, which can help in predicting the effectiveness of particular therapies.

With the new, JAMA Ophthalmology paper, the researchers showed that the approach can also measure severity scores over time, and thus monitor the progression of ROP. Retrospectively analyzing images from 5255 clinical examinations of 871 premature infants who met the ROP screening criteria – and now using a four-tier grading classification system – they found that the median severity scores in the eyes that eventually required treatment were higher than in those that did not. This suggested a correlation between the scores at given points in time and clinical progression of ROP in premature infants. Ultimately, the researchers wrote, the deep learning approach could have prognostic value in identifying infants who are likely to develop ROP requiring treatment.

“Advances in machine learning – specifically, deep learning – have great potential in all aspects of medical imaging,” said Jayashree Kalpathy-Cramer, director of the Quantitative Translational Imaging in Medicine (QTIM) group at the Martinos Center and co-senior author of the JAMA Ophthalmology paper. “In ROP, we are really excited to develop a tool that can quantify disease severity based on imaging that can then be used to assess whether the babies are progressing and/or responding to therapy. We hope that this is a small step forward in reducing preventable blindness in the US and worldwide. We are optimistic that the dissemination of such technologies can also improve access to care.”