A dental technician captures an X-ray, the patient leaves the chair, and the clinic assumes the image is sufficient for a claim. Three days later, the insurance provider rejects the reimbursement because the image is slightly blurred or improperly aligned. The clinic loses revenue, and the patient must be called back for a second visit, creating a friction point that damages the patient experience and the practice's bottom line. This operational gap is a systemic issue in dentistry, where up to 20% of insurance claims are initially denied due to missing or low-quality imagery. For years, the only solution was manual retrospective review, which happened long after the patient had exited the building.
Scaling Real-Time Quality Assurance to 10,000 Sites
Henry Schein One addressed this inefficiency by developing Image Verify, an AI-driven solution that transforms X-ray quality control from a retrospective chore into a real-time feedback loop. Rather than diagnosing pathology, Image Verify assigns a quality score from 1 to 5 immediately after an image is captured. This allows technicians to decide on a retake while the patient is still in the clinic, directly attacking the 20% insurance rejection rate.
The trajectory of the product's deployment highlights a massive surge in clinical adoption. Moving from a concept design phase in the fall of 2025 to full production in just a few months, the service began with an initial rollout to 250 locations. By the end of April 2026, the footprint expanded to over 10,000 active sites, representing a 43-fold growth in adoption. To date, the system has processed more than 11 million X-rays, with current throughput averaging 1.5 million images per week.
To maintain this scale, the architecture relies on a sophisticated orchestration layer. The system utilizes Amazon Elastic Kubernetes Service (Amazon EKS) to manage the application layer and coordinate requests coming from practice management software. These requests are then routed to machine learning inference endpoints powered by Amazon SageMaker AI. By decoupling the traffic management (EKS) from the heavy computational lifting (SageMaker), the system ensures that the high-volume stream of dental images does not bottleneck the user interface. The result is a clinical response time that feels instantaneous: a median round-trip latency of 1.4 seconds. Even at the P90 mark, where the slowest 10% of requests are measured, the latency remains at 2.2 seconds, ensuring that nearly every technician receives a score in under three seconds. Throughout millions of inferences, the system has maintained a remarkably low error rate of 0.01%.
The GPU Pivot and the Regulatory Shortcut
Achieving sub-two-second latency across thousands of global sites required more than just powerful hardware; it required a fundamental shift in how data moves through the pipeline. During the initial scaling phase, the team encountered a critical performance wall. Profiling revealed that the preprocessing pipeline—specifically image decoding, normalization, and resizing—was running exclusively on the CPU. This created a severe bottleneck where the CPU reached maximum utilization while the expensive GPU resources remained largely idle. In many large-scale ML systems, this CPU saturation masks the available GPU headroom, leading engineers to mistakenly add more instances, which increases costs without solving the underlying latency.
Henry Schein One resolved this by migrating the entire preprocessing stage directly to the GPU. By handling data transformation on the same hardware used for inference, they eliminated the CPU bottleneck and significantly increased the throughput per instance. This was paired with a transition to an asynchronous inference pipeline, allowing the system to handle multiple requests without waiting for each individual process to complete. To ensure stability, the team implemented an A/B testing framework and a zero-downtime deployment pattern, iterating through more than 60 detailed optimizations before the wide-scale rollout to 10,000 sites. This lean GPU fleet allows the company to support massive expansion without a linear increase in infrastructure spend.
Beyond the hardware, the most significant "twist" in the Image Verify strategy was the conceptual definition of the product. In the medical AI space, most companies race to build diagnostic solutions that identify diseases. However, diagnostic AI is subject to grueling regulatory scrutiny and slow approval cycles that can freeze product iterations for months or years. Henry Schein One intentionally defined Image Verify as a Quality Solution rather than a Diagnostic Solution. The AI does not look for cavities or bone loss; it only determines if the image is clinically usable.
This distinction created a regulatory bypass. Because the tool assists with technical execution rather than medical diagnosis, it avoided the most restrictive medical device hurdles. This allowed the development team to employ a lean, iterative approach, updating the model based on real-world feedback in days rather than years. This strategic positioning not only accelerated market entry but also turned the tool into an educational asset. New technicians now use the 1-5 scoring system as a gamified training tool, competing to improve their technical skill and raising the overall standard of care across the network.
By separating the technical requirement of image quality from the medical requirement of diagnosis, Henry Schein One transformed a regulatory obstacle into a competitive advantage. The combination of GPU-accelerated preprocessing and strategic product positioning proves that in medical AI, the path to scale is often found by solving the operational friction surrounding the clinic rather than the clinical friction of the diagnosis.
