1. There is a difference between a model that performs well on a benchmark dataset and one that can be trusted to inform a clinical decision. Where does that gap show up and what did you have to change about how you were measuring performance?
In radiology AI, standard benchmark datasets typically measure how well an algorithm matches a human radiologist’s annotations. For prostate cancer, this has typically been done by measuring the overlap (via Dice score) between AI output and an MRI region of interest (ROI). The gap shows up immediately in the clinic because human radiologist MRI annotations are a flawed benchmark. Radiologists can only see what is visible, and standard MRI routinely misses the full spatial extent of prostate cancer, underestimating tumor diameter by 11 millimeters on average. If your benchmark rewards an AI for perfectly mimicking a flawed human MRI contour, you are automating under-diagnosis.
For AI to go beyond the expert, you need the real ground truth: surgical pathology. We developed our AI with the best ground truth available. We created custom 3D molds derived from the pre-operative MRI to slice radical prostatectomy specimens in exact alignment with imaging planes. This allowed us to map the real tumors back onto MRI with millimeter accuracy, allowing our models to train against the actual tissue pathology, rather than MRI shadows. This unique approach is what allowed us to increase physician sensitivity to tumor extent from 37% to 97% and improved surgical margin accuracy from 67.2% to 84.7%.
2. Clinical validation requires demonstrating that a model behaves across patient populations, imaging conditions, and clinical contexts that weren’t represented in the training data. What did that process reveal about your model that internal testing hadn’t?
It is well reported that AI models trained on pristine datasets from academic medical centers often fall flat during real world use. However, while imaging alone is brittle, multimodal data fusion is resilient. In addition to MRI, our model uses the real pathology samples obtained during the diagnostic biopsy, tracked in 3D. This grounds our model in a pathological reality, significantly reducing the subjectivity inherent in human generated labels and eliminates diagnostic hallucinations. However, when we took our model to the real world we found that many urologists were not doing the 3D tracking of their biopsy samples that we expected. Prior to Unfold AI there was no downstream system to use that information, so many didn’t bother. However once they learn about Unfold AI we have found many urologists quickly change their behavior – often before they have our software.
3. Surgeons interacting with the system during development and clinical evaluation gave you feedback that a dataset cannot. What did that feedback change about the product not at the interface level, but at the level of what the model was doing?
Once our product was in front of surgeons, they all had the same question – does our AI model’s cancer estimation on the prostate boundary correlate with extracapsular extension (ECE) beyond the boundary? And we are glad they asked! Subsequent studies at UCLA and Stanford found that Unfold AI reduced false negatives for ECE by 40% to 60%. Also, accuracy in predicting seminal vesicle invasion (SVI), or the spread of cancer outside the seminal vesicle, rose from 52 to 92% over expert radiologists. This has huge implications for nerve-spare prostatectomy and radiation treatment planning. Furthermore, for surgeons pursuing focal therapy there is a need to know the likelihood a margin will include the full tumor. To meet this need, we developed a metric called the Encapsulation Confidence Score (ECS), which calculates the exact statistical likelihood (R2 = 0.98) that a user-selected margin boundary will successfully encapsulate all clinically significant disease. More recently, surgeons asked for dynamic outputs for a new Surgery Mode and custom platform integrations for robotic consoles like Focal One HIFU and Da Vinci systems. These allow the model’s spatial probabilities to guide real-time tissue dissection and nerve-sparing decisions.
4. The FDA process surfaces assumptions that were invisible during development about how the tool will be used, by whom, in what clinical context, with what level of supervision. What assumptions did the regulatory process force you to make explicit?
There was initially a disconnect between how the FDA traditionally evaluates prostate radiology AI (focusing on diagnostic lesion scoring using human labels) and our approach, which relies on post-surgical pathology ground truth to optimize treatment planning. To bridge this gap, regulatory review spurred us to conduct a 10-physician multi-reader study comparing clinician performance with and without Unfold AI. The study proved that while radiologists and urologists are effective at lesion detection, they struggle with 3D margin definition on their own, but achieve significant accuracy gains when assisted by our software. This demonstrated to regulators that our intended user base appropriately includes both urologists and radiologists, establishing Unfold AI as a true treatment-planning tool rather than a standard diagnostic aid.
5. A surgeon who has spent decades developing judgment about treatment planning is being asked to incorporate an output from a system they didn’t build and may not understand. What did you learn about how that trust develops and what approaches to building it turned out to be wrong?
We initially learned that pushing a standalone margin metric onto surgeons felt disjointed from their clinical reasoning, so we evolved our output into an interactive 3D Cancer Estimation Map (CEM). Seeing the 3D heatmap naturally prompted surgeons to ECE risk, which we validated through UCLA research led by Dr. Wayne Brisbane to quantify how our AI correlates directly with capsular disease spread. By answering the exact clinical questions surgeons were already asking, we met them directly within their existing treatment decision-making framework. With our new intraoperative Surgery Mode, clinicians can now seamlessly marry these predictive metrics directly to the live surgical scene on their console.
6. The physicians who adopted the system earliest are different from the ones who came later. What distinguished the early adopters and how did that difference change about how you approached broader adoption?
Our early adopters were academic surgeon-scientists and clinical researchers at UCLA and Stanford who were actively leading clinical trials in focal therapy and robotic surgery and wanted to push the science forward.
As we moved toward broader adoption among high-volume community urologists and private regional practices, like Arizona Urology and U.S. Urology Partners, the priorities changed. Community physicians don’t have time for complex academic software setups. They care about workflow friction, patient satisfaction, and financial sustainability. That difference led us to pivot our broader adoption strategy from academic trial presentations to commercial viability – ensuring our technology was reimbursable in the real-world. We secured a dedicated AMA Category III CPT billing code, established a national Medicare payment rate, and most recently earned inclusion in the regional Physician Fee Schedule for the west coast.
7. Hospital procurement is slow, complex, and involves stakeholders across clinical, IT, legal, and finance functions who each evaluate the technology differently. What did you learn about navigating that process?
Satisfying clinical validation, FDA clearance, and IT requirements, including our SOC 2 Type II certification and EHR playbooks, solved technical friction long ago. Procurement velocity, however, really depends on a strong physician champion who can navigate internal site politics and prioritize review. Equally important is a clear, established payment message, which directly aligns hospital finance with clinical adoption. Because each health system evaluates technology differently across clinical, IT, legal, and financial stakeholders, we continuously raise the tide by expanding our peer-reviewed clinical and economic evidence. As regional adoption expands and neighboring sites activate Unfold AI, competitive momentum creates a powerful catalyst that accelerates institution-wide approval.
8. Integration with existing clinical workflows and hospital IT infrastructure is where many clinical AI tools fail in deployment. What did real-world integration reveal about the assumptions the system was built on?
In an academic development lab, you implicitly assume that input data arrives neatly packaged in standardized formats. Hospitals don’t work that way. We found that DICOM metadata varies wildly across MRI scanner vendors, fusion biopsy systems store spatial coordinates in unstandardized file types, pathology reports arrive as unstructured PDF text.
Real-world integration revealed that our ingestion architecture had to be far more flexible than our core algorithm. We had to build a robust multi-modal data normalization pipeline that could ingest raw, heterogeneous data streams from multiple PACS and EHR systems. It had to extract key variables and register them into a unified 3D coordinate space in real time. At high volume sites manual data entry is a workflow killer, especially for DICOM images and novel fusion biopsy data formats. Most urology staff live out of the EHR, so asking them to do unfamiliar manual work had a high error rate. So ingestion had to be fast and automatic without adding manual data-entry overhead, especially for DICOM images and the fusion biopsy data.
9. A platform that began as a research project in an academic medical center and became a clinical tool used in real patient care has gone through a transformation. What is the decision that you now recognize as the one that made the difference?
Two of our founders, Shyam Natarajan and Leonard Marks, understood the untapped potential of multimodal AI long before it was mainstream. That novel approach is what got us an FDA breakthrough designation, our novel CPT code in augmentative AI, and a huge boost in retrospective studies to justify the CPT code and subsequent payment. Importantly, it earned the patience of our early adopters while we figured out the complex process of selling to hospitals. No one would wait so long for a “me too” product; we had to be new and shiny but also impactful.
The traditional medtech playbook is linear: build an algorithm, publish academic papers, pursue FDA clearance, and only then (often five to six years later) figure out how hospitals will pay for it. We recognized early that an un-reimbursed clinical AI tool dies in the lab regardless of its accuracy. By training our models against whole-mount 3D pathology molds while simultaneously designing our software for FDA 510(k) compliance and engaging the AMA and CMS early for Category III CPT coding, we cleared the commercial valley of death in just 18 months and under $10 million in capital. That decision turned a small academic research project into a sustainably deployable, reimbursable platform actively improving patient care across the country.
Author Bio:
Joshua Shubert joined Avenda Health as their first Software Engineer in 2018. Now as CTO, he is managing a team of skilled software engineers and scientists to advance their AI prostate cancer mapping technology, expanding access to physicians and patients across the country. Joshua holds a M.S. in Electrical Engineering from Johns Hopkins University and has gained industry experience in cloud-based medical software, as well as developing a strong background in deep learning, medical imaging and robotics.
