In September 2016, four colleagues and I published a chapter in Lecture Notes in Computer Science, alongside the first International Workshop on Deep Learning in Medical Image Analysis at MICCAI. It was twelve pages long. It described a fully automated pipeline for applying Graf's method to hip ultrasound scans in infants, using a deep convolutional network with an adversarial component to segment the ilium and acetabular roof, and a set of post-processing rules to compute the alpha angle from those segmentations. On a bronze-standard test set, the pipeline produced alpha angles that fell within 5 degrees of the expert reading in 77% of images. Assuming the expert as ground truth, we reported no false negatives and a 14% false positive rate, close to the disagreement rates already documented between human experts. In three cases where the machine and the human differed by more than 10 degrees, our own senior review concluded that the machine was right.
Ten years on, I want to look at what the paper predicted, what happened to the five authors, and what we could not have predicted.
What the paper argued
The technical case was that Graf's method could be automated end to end from the raw ultrasound image, and that the automation could match or exceed expert-level agreement. That case has held. The pipeline we described has been reproduced and improved on by multiple academic groups in the years since. As of 2026, at least three FDA-cleared AI products for DDH ultrasound assessment are on the market, and the alpha-angle problem is regarded as solved to within inter-rater variability, which is the ceiling anyone can realistically ask for on a subjective measurement.
The clinical case underneath the technical one was subtler and, in my view now, more important. We argued that Graf's method was so operator-dependent, requiring years of training to reach reproducible performance, that its clinical use was structurally limited. In every country practising universal ultrasound screening for DDH (Germany, Austria, Switzerland, Israel), the value of the programme rested on a scarce and expensive workforce of expert sonographers. In every country practising selective screening (the UK, the USA, Canada and most of the rest of the world) the case against universal screening rested largely on the same workforce problem. Take the workforce problem away, we argued in effect, and the calculation shifts. The paper's implicit prediction, though we never used the word, was that AI would enable universal DDH screening to become the standard of care in more countries than were doing it in 2016.
That prediction has not come true. And the reasons why are the subject of a separate essay I am editing at the moment, called The Evidence Constraint. But the DDH story is the cleanest example I know of the mechanism the essay describes, so let me lay out the details.
What happened to the five of us
The 2016 paper had an unusual author line for its time. David Golan was in the Stanford statistics department. Yoni Donner was in Stanford computer science. Chris Mansi was in the Stanford Graduate School of Business. Jacob Jaremko was a paediatric and musculoskeletal radiologist at the University of Alberta, one of the very few clinical academics at that moment who could read both a Graf ultrasound and a Keras training script. And I was a paediatric orthopaedic surgeon at the Royal London Hospital, providing the clinical need that got the collaboration started.
Around the time our chapter appeared, Mansi, Golan and I co-founded Viz.ai. Donner was one of the early hires. The company pivoted deep learning to a different clinical problem, large-vessel occlusion detection on brain CT for acute stroke. Viz.ai went on to raise several hundred million dollars, become the first AI company to secure a CMS New Technology Add-on Payment, and become the standard AI platform for stroke pathway acceleration in the United States. The DDH chapter was the technical proof of concept for what a small team could build in medical imaging AI in 2016. The commercial company was always going to be built around a clinical problem with a cleaner economic pathway.
Jaremko took the DDH work directly commercial. In 2018, he co-founded Medo.ai at the University of Alberta with Dornoosh Zonoobi (a former Alberta Innovates postdoc who became Medo's CEO) and Jeevesh Kapur (a radiologist from Singapore who was in the CUDL group with us on the 2016 paper). Medo built exactly the workflow our chapter had described in principle: a handheld ultrasound acquisition guided by AI, so that a non-expert could produce a diagnostic-quality scan and get an automated read on the spot. Medo Hip received FDA clearance in 2020, becoming the first FDA-cleared AI product for DDH. In July 2022, Exo Imaging, a Redwood City handheld-ultrasound company with over 320 million dollars of funding, acquired Medo and integrated its Sweep AI into the Exo point-of-care platform. Jaremko stayed at the University of Alberta, where he now holds an endowed chair in diagnostic imaging and a CIFAR Canada AI Chair. He has spent the years since the Exo acquisition arguing, with Alberta-specific numbers, that AI-enabled universal DDH screening is now the right policy call. Roughly four DDH babies are born per day in Alberta. Roughly two adult hip replacements are performed per day at least partly attributable to missed or delayed DDH.
I kept both careers going. I have been a Viz.ai co-founder and I have been a consultant paediatric orthopaedic surgeon throughout. My paediatric clinic at Barts Health still sees DDH babies weekly. In the UK, we still screen selectively, using the Newborn and Infant Physical Examination with risk-based ultrasound as a second line, and the referral criteria I work to are essentially the ones I inherited when I became a consultant. What has changed is that I know what a fully automated pipeline can do with a Graf scan, and I also know that in most of the Anglophone world, the standard clinical answer to that fact is still a shrug.
What we could not have seen from 2016
Two things.
The first is that the DDH problem, in retrospect, was the wrong problem to build a company on if your goal was to change medicine at scale. It is a screening problem, not an emergency. It plays out over decades in the form of adult hip disease, which is billed to a payer thirty years downstream from the child you might have caught. Nobody's budget is defended by preventing a case of hip osteoarthritis in 2055. Our decision to build Viz.ai around stroke rather than DDH made complete commercial sense the moment we asked which clinical problem had a fast payer, a measurable acute outcome and a workflow already digitised end to end. Stroke has all three. DDH has none.
The second is that even a technically solved problem in medical AI can sit for years without changing standard of care, and there is nothing about the strength of the evidence that determines when or whether the change comes. The 2016 paper and the products that followed did what they said they would do. The evidence for AI-enabled DDH ultrasound is not the missing piece. What is missing is a whole apparatus of health-economic modelling, screening-trial design, professional-society endorsement, commissioning pathway and pricing. This apparatus moves slowly. A technically correct paper in 2016 has still not, in any jurisdiction that was practising selective screening then, altered its DDH screening policy in 2026.
The one point I would carry forward
The DDH story, taken alone, is an accurate small-scale rehearsal of the situation the entire field of clinical AI now finds itself in. Model performance in DDH is no longer where the value is created. It has not been for some years. The value is created downstream, in the evidence infrastructure that supports a policy change and in the workflow depth that puts the automated read in front of the person who can act on it.
I am proud of the 2016 paper. It was a careful early piece of engineering on a clinical problem I care about, and the pipeline it described works and has been reproduced and cleared for use. It just turns out that the pipeline was the easier part.
For now, the DDH babies in my hospital still see a multidisciplinary team trained to the highest standards. The DDH babies in Alberta will, I hope, one day be screened by an ultrasound probe held by a family doctor with an AI reading quietly in the background. That day is closer than it was in 2016.