From CRADLE to Care: Your Photos Can Help Scientists Build an App for Parents to Diagnose Serious Child Eye Disease Early.
Monday December 9, 2019
A White pupil is the most common early sign of eye cancer in young children, and it’s usually seen first in photos. Do you want to turn your smartphone into a powerful cancer-detector? Mark Billings, WE C Hope USA Director, tells the story behind the CRADLE app and explains how you can help scientists improve it.
White Eye – A Critical Warning Sign
At one time or another, most parents who have a child with retinoblastoma have seen a whitish reflection in the pupil of the affected eye, observing it either in photographs or in dim light with their naked eye. The reflection is variously described as “a glow”, “creamy”, “milky”, “pearly”, “transparent”, “like a crescent moon in her eye”, or “like a cat’s eye at night”. Though most commonly described as white, the reflex can range from bright redish-white, orange-white, yellow or grey. Colloquially, it is known as “the cat’s eye reflex”, but the medical term is leukocoria.
When we take a photo with a flash, we usually see a red reflection in the eye. This reflection is a good indication that the eye is functioning normally – the red glow is caused by healthy pigment epithelium and vascular layers supporting the retina, seen through the pupil – which expands momentarily in the flashlight. If a white reflex is seen, it means something is preventing the light from reaching the retina. The beam of light is then reflected back out of the eye, creating the appearance of a white glow in the pupil. Depending on the location of the problem in the eye, and the angle at which the light enters the eye, the white reflex may not appear in every photograph.
A white eye can be one among many symptoms of various eye disorders, the most serious of which is retinoblastoma, an aggressive eye cancer affecting babies and young children. White eye reflex is considered the cardinal symptom of retinoblastoma. The American Academy of Pediatrics, and similar bodies worldwide, recommends all children showing signs of white eye be urgently examined by an ophthalmologist.
Red Reflex Screening and Camera Detection
General Practitioners, pediatricians and opticians may use the red reflex test to look for signs of eye disorders during a child’s routine wellness visit. This non-invasive exam uses handheld torch to shine a light into the pupil of each eye to check for proper eye alignment and healthy red eye reflection. However, this simple routine check is often not done, or is performed incorrectly, and fails to spot white eye.
Red reflex office exams will only visualize about 2% of the retina at the back of the eye. However, the retina lines much of the inner eye, from the back of the eye to the border with the iris (the colored ring) at the front. As the child grows older, new tumors arise from retina towards the front of the eye and will likely only be detected on routine exam when they are already fairly large. The pupil needs to be dilated, and the eye considerably manipulated to allow a clear view of the entire retina.
Many parents see “white eye” in photographs before doctors observe it in person. They may try to take photos with the specific goal of capturing the white glow they have seen.
Repeat appearance of the strange white glow in just one child in family snaps often becomes the motivation to seek medical care. Some parents report looking back through their pictures, after diagnosis, to discover that photographs with “white eye” had been taken months or years before diagnosis.
That’s what happened to Elizabeth and Bryan Shaw, and their son Noah.
Bryan explains “His mum noticed white eye in his pictures and told the pediatrician. He was diagnosed later that afternoon by an ophthalmologist but up to that point he had passed all of his red reflex tests. We went back and started looking through all of our baby pictures and we noticed that white eye emerged in pictures taken at 12 days old. If we had caught the cancer then, he wouldn’t have lost his eye or had his other eye radiated. His vision would be much better because the tumors would have been smaller when they were discovered.”
Noah, aged 7 months, 3 months after diagnosis. (Credit, Bryan Shaw)
An App for That!
Bryan is a scientist – he thrives on defining a question, and researching, developing, testing and refining a response. So when he learned about leukocoria and its key role in early diagnosis, he began exploring whether technology could help parents detect potentially blinding and lethal eye disease early – when the reflex first appears in photographs.
“If I had some software telling me, ‘Hey, go get this checked out,’ that would have sped up my son’s diagnosis and the tumors would have been just a little bit smaller when we got to them. There might have been fewer. Maybe we wouldn’t have had to remove Noah’s right eye to prevent the cancer spreading.”
As Associate Professor in the Department of Chemistry and Biochemistry at Baylor University in Waco Texas, Bryan reached out to Baylor computer scientist Greg Hamerly to explore their options. Inspired by the extensive photo-library documenting Noah’s first months of life, the pair created plans to build a photo-analysis smartphone app. They called it “White Eye Detector”.
Greg’s research team developed a program that uses a form of artificial intelligence called machine learning to detect leukocoria in photographs. Machine learning trains the system to complete a task by example. To successfully train the app to recognize white eye, the developers needed to feed the program thousands of photographs showing white eye in children with confirmed eye conditions. They began with Elizabeth’s photo-library which contained more than 10,000 unique moments from Noah’s first 3 months of life.
Bryan quickly discovered this volume of photography is not unique. He received donations of unsorted photo-libraries from parents of children with retinoblastoma and other eye conditions – they heard and read early media reports about the app and were keen to help.
The ease of digital technology today means that modern parents take hundreds and thousands of photographs documenting every aspect of their children’s lives. The donations proved to Bryan, Greg and their team that the app could become a powerful addition to occasional routine red-reflex screening – screening that is still totally absent for millions of children worldwide. Every photo represents potential for a single red reflex test that can identify leukocoria early.
The app scans through unsorted libraries on a smartphone or tablet, and highlights any that contain potential leukocoria. Although it scans automatically on Android devices, it requires activation before every use on IOS. When a photo of concern is detected, the app, suggests the child be taken to a doctor for examination.
In addition, a screening mode enables the user to shine the device’s torch light into a child’s eyes, and use the camera scan for white reflex. An image of the child appears on screen, each eye surrounded by a red (abnormal) or green (normal) results box.
CRADLE detection of leukocoria caused by cataract (bottom right) and retinoblastoma (all other panels). Red box indicates positive leukocoria detection; green box indicates negative leukocoria detection. The inset in the upper right corner of each photo shows a magnified view of the leukocoric pupil detected by CRADLE. (Credit Munson et al., Sci. Adv. 02 Oct 2019: Vol. 5, no. 10.)
To protect privacy of the user and those in their photos, the app does not require that images be uploaded to a server, and it does not track user activities. Software is downloaded, to scan directly on the device. If leukocoria is detected, the user is invited to upload the concerning photos to aid research, and give consent at that point.
Bryan demonstrated the app before the U.S. House of Representatives Subcommittee on Research and Technology. He recalls during the testimony, I showed a video demonstration of the CRADLE app on Noah. When the images and video began to be shown, there was an audible bipartisan gasp from the congress. Noah was sitting behind me during the testimony and can be seen throughout the video, playing with crayons.”
Bryan’s testimony runs from 27:34 to 33:38.
A Free Tool for Parents
The first iteration of the app was released in 2014. The developers felt strongly that, for humanitarian reasons, their prototype should be immediately available at no cost to parents worldwide. Since then, it has been downloaded to more than 200,000 Android and iOS devices on every continent excluding Antarctica.
Now renamed CRADLE – ComputeR Assisted Detector of LEukocoia, the app can be downloaded free from the App Store and Google Play under the name “White Eye Detector”. The technology remains free of charge to all users.
Bryan noted ‘We can’t track and quantify the impact of these CRADLE prototypes. However, media outlets have reported incidents of parents who initiated early diagnoses of retinoblastoma and other conditions in their children, after using the app on their children’s photos and showing the results to their doctor.”
The first success stories were from two different families in Germany, who used CRADLE to detect leukocoria and triggered early diagnosis of retinoblastoma in their children. Since then, anecdotal stories have trickled in of parents successfully using the app to catch leukocoria and aid diagnosis of different eye disorders including Coats’ disease, cataract, and myelin retinal nerve fiber layer.
Examples of leukocoria detected in childhood photos by the CRADLE app. (A) Pathologic leukocoria is caused by reflection of light off abnormal ocular surfaces in conditions like cataract, Coats’ disease, and retinoblastoma, or abnormal reflection of light off the optic disc, such as in refractive error, anisometropia or strabismus. (B to F) Examples of pathologic leukocoria in photographs of test children from the CRADLE study who have been diagnosed with (B) hyperopia, (C) retinoblastoma, (D) Coats’ disease, (E) anisometropic amblyopia, and (F) cataract. (G) Child without an eye disorder exhibiting normal red reflex (right eye) and physiologic leukocoria (left eye) caused by reflection of light off the optic disc during off-axis photography. Insets show the leukocoric pupil detected by the CRADLE app, operating on iPhone X. (Credit Munson et al., Sci. Adv. 02 Oct 2019: Vol. 5, no. 10.)
How Good Is The App?
Results from the first real-world test of the CRADLE app were published in October in the journal Science Advances. They reviewed Nearly 53,000 childhood photographs from 40 children, 20 with different eye disorders and 20 healthy controls. The app correctly detected leukocoria in 16 of 20 children (80 percent) with diagnosed eye conditions. On average, leukocoria was identified in photos taken 1.3 years before the child was diagnosed.
This is a highly influential tool, potentially giving parents an early alert that something might be amiss with their child. Retinoblastoma can begin to spread from the eye to the brain or bone marrow as soon as 6 months after leukocoria becomes visible to the naked eye. So this early alert really can save lives and reduce the burden of treatment needed to cure the child and potentially save sight.
The app did not detect every photo with white eye. This is known as a “false negative”. Rather it identified about 1/3 of images with leukocoria when it was present. It was found to be more reliable when scanning a large library of everyday family photos, than a single photo or a small collection of posed images. The more photos taken in a variety of environments, the more opportunity there is for light to reflect off the ocular lesions causing white reflex, regardless of their location in the eye.
Through constant training, CRADLE’s algorithm has grown ever-more adept and is now able to identify very slight instances of white reflex. Bryan was surprised at the low levels of detected leukocoria highlighted in the research. “This is one of the most critical parts of building the app. We wanted to be able to detect all hues and intensities of leukocoria. As a parent of a child with retinoblastoma, I am especially interested in detecting the traces of leukocoria that appear as a ‘grey’ pupil and are difficult to detect with the naked eye.”
Research also demonstrated the app’s skill in detecting white eye even when the photo has a wide field of view, and the child’s pupil is very hard to see and assess with the naked eye. These wide-view photos are common in family gatherings, play and experiential settings throughout early childhood.
Examples of Apathologic leukocoria detection by CRADLE in photographs with a wide angle and low pupil resolution. Bottom insets show magnified view of the leukocoric pupil detected by CRADLE (red box). In left image, the child’s left leukocoric pupil (red box; cropped pupil = ~ 111 pixels) is affected by retinoblastoma. In right image, the child’s left leukocoric pupil (red box; cropped pupil = ~ 18 pixels) is associated with hyperopia. Bottom insets also lists the percentage of the original photograph comprised of the cropped pupil. (Credit Munson et al., Sci. Adv. 02 Oct 2019: Vol. 5, no. 10.)
Leukocoria was reported in less than 1% of photographs where no eye condition was present. While this very low “false positive” exceedingly encouraging, it is a cause for concern and a focus for further research. Even though false positive reports occur infrequently, a 1% rate could result in worry for tens of thousands of parents, and unnecessary eye exams for children in a populous country like the United States.
The research team are keen to train CRADLE further, to reduce the rate of false positives substantially below 1%.
A key challenge is that the app cannot currently distinguish between “pathologic leukocoria” – white eye caused by an eye condition, and “physiologic leukocoria” – white eye caused by light reflecting off the optic nerve head in a healthy eye, where there is no underlying condition.
Photos of children with eye conditions show pathologic leukocoria 10 times more often than photos show physiologic leukocoria in children with healthy eyes. But Bryan explains the app can’t yet distinguish the two – “they often look identical.”
The app invites users to upload any identified leukocoria photos to Baylor’s database, so they can be used to continue refining the algorithm. Through continued training, Bryan and his dedicated team expect the app will become more skilled at differentiating between the two reflexes, and increasingly eliminate false positives that may cause needless worry for parents, children and their wider family.
Despite the obvious challenge of false positives, this technology is an excellent resource for both families and healthcare professionals alike. CRADLE is poised to revolutionize the way parents and professionals identify eye cancer and other blinding childhood conditions early, increasing potential for prompt access to life and sight saving care.
Parents, Pediatricians and Pictures
CRADLE’s impact and potential stems from the booming supply of smartphones. In the past five years, the reported ownership of smartphones has skyrocketed in most nations. According to the Pew Research Institute, ~ 92% of adults in the United States owned a smartphone in 2015 – no big surprise there. But a comparison of the numbers of parents versus doctors might surprise you.
There are around 37 million parents in the U.S. with children under 5 years old. This population dwarfs the 27,000 general pediatricians and 123,000 family and general practitioners in the U.S. who can examine a child’s eye during a routine exam. This pattern is replicated in countries around the world – in many developing countries, the gap between smartphone ownership and physician numbers is significantly larger.
The effectiveness of traditional red-reflex screenings during a general physical exam is limited. 50-90% of parents report observing leukocoria when their child is diagnosed with retinoblastoma. However, only 8% of children are diagnosed due to detection of leukocoria in paediatrician and primary health care conducted eye exams.
In contrast, CRADLE detected leukocoria in 80% of children age two and younger with confirmed eye conditions. That 80% threshold is regarded by ophthalmologists as the ‘gold standard’ of sensitivity for similar devices.
Clearly, the app is a powerful tool, empowering parents to effectively supplement clinical screening, or to screen their children when conventional screening is not provided. It is useful for children of all ages, but most valuable for very young and non-verbal children who can’t explain that their vision is impaired. Pre-verbal children are the age group at highest risk of leukocoria-causing conditions like retinoblastoma and congenital cataract. The more photos parents take of their children, the more opportunity they have to find telltale early signs of a developing eye issue.
Bryan is certain the app can help prevent children from going blind and dying. “A parent, and their camera, are the first line of defense in screening and preventing these deaths. Every parent should have an app for leukocoria detection. So mom and dad, take lots of pictures, and if there is a problem, this app will probably catch it.”
The CRADLE team emphasize that the app isn’t yet FDA-approved and is not intended to be a substitute for a clinical diagnosis. An ophthalmologist will need to carefully examine both eyes to diagnose a specific condition causing the white eye. However, the app gives parents the knowledge and evidence to seek medical care in the first place, and that is the vital first step towards diagnosis and treatment.
Example of early detection of leukocoria by CRADLE in a child with diagnosed retinoblastoma. (Credit Munson et al., Sci. Adv. 02 Oct 2019: Vol. 5, no. 10.)
CRADLE has significant potential to improve the outcomes of children with sight and life threatening eye conditions in remote communities and under-resourced countries around the world.
In countries with relatively small populations and vast geography, such as Canada and Australia, families living outside urban settlements may travel significant distances to access regular health care. As a result, children living in rural communities may have fewer routine health checks than those living in towns and cities.
In developing countries across the globe, the number doctors per 10,000 people is significantly lower than in developed countries, which contributes to late diagnoses of childhood eye conditions. While more than 95% of children with retinoblastoma are cured in developed countries, global survival is barely 30%.
In Kenya for example, physician density is lower than the U.S. by a factor of 10, However, in 2015, Pew reported that smartphone ownership among Kenyans was 34% and rising among 18-34 year olds.
Parents, their smartphones and CRADLE could help fill the gaps created by lack of access to routine screening for millions of children around the world. An app is no substitute for a doctor and a thorough eye exam, but it is definitely better than nothing.
CRADLE gives parents clear evidence that their child needs a professional eye exam, a test they may not otherwise seek due to distance and logistics, or lack of resources. That evidence is easily conveyed to the medical professional from whom they seek care. So the technology, even at its prototype stage, may dramatically speed up a child’s diagnosis. Especially in situations where a child’s eyes are not routinely screened, and where primary doctors are unfamiliar with the early warning signs.
Doctors are interested in using the app as well to aid their own screening process when children first arrive at a medical facility. In a busy clinic with few staff and diagnostic tools, initial screening with the app can help decision-makers identify which children urgently need care, and prioritize limited resources
A Deeply Personal Mission
Studying childhood eye disorders, especially retinal cancer, is a challenging experience for Bryan. His son is now 11 years old and cancer-free, but he walks with Noah every day through the life-long and life-changing effects of bilateral retinoblastoma.
“Examining family photographs of children with eye cancer can be heartbreaking. You see the innocent, beautiful child and their loved ones, maybe a first birthday or being held by grandma, or learning to eat. Everyone is happy. But you see the strange-looking pupil in the child or infant, and know it is caused by an aggressive cancer that no one in the photograph knows about.
“I’ve been in those pictures, I’ve taken those pictures. I’ve mourned over those pictures, after the late diagnosis.
“During the course of this study, we lost one child in the cohort; Max Gottstein, and we dedicated the paper to him. A research fund was established in his name at Children’s Hospital of Philadelphia (We are not affiliated with this research fund). Max’s father is a big emotional supporter of ours, and helped get the CRADLE app advertised on Verizon’s homepage.”
You Can Help Train the App to Save Life and Sight!
Bryan and his team of scientists need help to improve CRADLE’s prototype technology. They are currently training the algorithm with an additional 100,000 photos, and exploring more features to reduce false positive reports.
To train CRADLE to better recognize pathologic white eye and dismiss physiologic white eye, Bryan needs even more photographs of our children. In particular, the team need photos of children from across Africa and Asia. That is the only way they can make the app more globally relevant, reliable and impactful to save children’s sight and lives in diverse populations.
Any leukocoric photograph of your child with a confirmed eye condition, and any leukocoric or non-leukocoric photograph of any healthy siblings, can be used by the researchers in two ways: training and testing.
Training the App
The more pictures that are fed into the machine learning algorithm, the smarter the app becomes in detecting pathologic leukocoria, and recognizing physiologic leukocoria as artefact in a healthy eye. For truly community-wide and global effectiveness, the app needs to be trained on photographs from a diverse set of children with different skin tone, eye color and shape, etc.
Testing the App
Researchers need large, unsorted libraries of photographs of children with retinoblastoma or another diagnosed eye condition whose symptoms include leukocoria. Using CRADLE, researchers retrospectively analyze these large libraries to assess if the app would have detected leukocoria before the child’s clinical diagnosis, and by what timescale. This testing is important to validate the app’s utility.
If you would like to help Bryan, Greg and their team of scientists continue to develop this powerful free technology, please email Dr. Bryan F. Shaw directly at email@example.com.
Noah, aged 3 months, 1 month before diagnosis. (Credit: Bryan Shaw)