The episode, “Navigating hospital politics, workflows, and entering the US market with Thomas Bertinotti,” features Thomas Bertinotti, CPO and COO of Therapixel, who has been building AI in medical imaging for roughly a decade. Thomas explains that Therapixel always had the US in mind - they designed the product using both US and French feedback from the start - but discovered that “US-ready” meant far more than model performance. The biggest gap was integration and workflow: US radiologists tolerate fewer clicks and expect information “at a glance,” which pushed Therapixel toward a highly integrated widget approach that fits directly into reading behavior rather than forcing extra steps.
Thomas contrasts what AI looked like 10 years ago versus today. Early models often produced only a study-level “yes/no” prediction without explaining where the finding was, which was a non-starter for real clinical use. Over time, the field moved from broad predictions to locating, characterizing, and guiding attention to specific lesions. He also describes a major shift from 2D mammography to 3D tomosynthesis in the US: Therapixel’s first FDA clearance (2019) was for 2D, but US demand rapidly centered on tomosynthesis, which required new training strategies, far more compute, and access to the right data. Hardware mattered, but Thomas says data access and the realities of clinical adoption were equally limiting - market studies might say “useful,” but true pull-through only happens once radiologists try the tool in practice.
A core insight in the episode is Therapixel’s focus on proving ROI in the US by measuring workflow impact inside the customer’s environment. Thomas describes running a baseline period where the widget is installed “in the background” to measure reading patterns, then turning on AI and showing a direct, site-specific improvement in reading speed. This only works because the widget approach gives visibility into workflow timing - something they couldn’t capture if the AI output lived only as static content inside the DICOM viewer. He also highlights a key France vs US workflow difference: in France, screening often includes a clinical exam (and sometimes ultrasound) alongside mammography, so faster image reading isn’t the bottleneck; in the US, reading throughput is much more central, making speed and efficiency a more compelling purchasing case than “quality” alone.
Thomas also gets into the realities of selling and “hospital politics.” In France, practices are smaller (often around 10 radiologists), decisions can be made by a small board, and a single clinical champion can carry a deal - though deal sizes are smaller. In the US, practices can be 60-80 radiologists (or large health systems), with dedicated IT, operational, financial, and security stakeholders involved. That makes selling more structured but also more complex: you need a champion and cross-functional alignment. Thomas notes that midsize practices can actually be easier than giant health systems because silos are smaller and stakeholders communicate more, while large systems can stall when clinical, IT, and operations don’t coordinate.
As Therapixel scaled, Thomas describes what “broke” first: speed and flexibility in shipping. Early on they could change priorities quickly, ship frequent releases, and even make late-night fixes because only a handful of sites depended on the system. With more customers and integrations, that became impossible. They had to standardize, plan releases months ahead, introduce staging and acceptance testing with customers, and reduce ad-hoc changes to avoid breaking production for many live sites. Thomas argues that building a real product organization early is critical in this category, and he credits Therapixel’s leadership with formalizing product management rather than assuming the CEO/CTO could “just do product” indefinitely.
Looking ahead, Thomas is skeptical that foundation models will magically erase the hard problems of radiology AI. He expects foundation models to help models do more tasks and potentially unlock viable solutions for rarer findings that were previously too costly to build, but he doesn’t think they’ve consistently outperformed carefully trained task-specific models yet. He also flags regulation as a looming scaling constraint: it’s hard to clear “100 findings” the same way the industry clears individual findings today. On lessons learned, Thomas emphasizes the hidden operational burden: beyond AI research, you need strong data engineering, longer AI release cycles, and especially AI Ops - the discipline of running models reliably in production across changing environments. Therapixel built dedicated AI Ops capabilities after seeing real-world processing failures, and this pushed reliability toward near-total processing coverage. He closes by predicting consolidation and/or increased vertical specialization (Therapixel’s focus is breast imaging), because the number of vendors is overwhelming for buyers and radiologists - and the market will inevitably simplify.
%20(1).png)
