At ECR 2026, a dedicated session titled “The Art of Ethical AI: Redefining Performance in Radiology” brought together experts to address the practical and regulatory dimensions of deploying AI in imaging departments. From changing European legal requirements to post-deployment monitoring, and the often-overlooked impact of AI on frontline staff, the session offered a grounded perspective on what responsible AI adoption actually requires.
No More Regulatory Hide and Seek
Dr. Hugh Harvey, a former radiologist and regulatory consultant from the UK, opened with an update on the EU AI Act. Medical AI devices remain classified as high risk, but the European Commission's November 2025 Digital Omnibus proposal moved medical devices from Annex 3 to Annex 1 of the Act. This means their high-risk status now derives from the Medical Device Regulation (MDR) rather than the AI Act directly. The practical consequence: the compliance deadline for vendors has been extended from August 2026 to August 2, 2028.
AI developers and distributors must provide electronic instructions for use, human oversight measures, and publicly available performance metrics. “No more regulatory hide and seek. I think we all want to see everybody playing fair, being honest about the performance of their systems, being honest about error rates and things like that. And I think the AI Act combined with the MDR should increase that transparency,” Harvey said.
Accountability extends equally to hospitals. Deployers must ensure appropriate use within the intended purpose, assign qualified human oversight, and maintain automatic logs of clinical events. "When you're purchasing a system, you are just as responsible for the oversight, the vigilance, the risk management, the monitoring of it. You cannot just rely entirely on the vendor." Non-compliance carries fines of up to 35 million euros for large companies.
Keeping Track: Post-Market Surveillance
Kicky Gerhilde Van Leeuwen, a healthcare AI specialist and researcher from the Netherlands, addressed the gap between pre-implementation validation and long-term performance monitoring. Most institutions invest heavily in retrospective analysis and acceptance testing but invest relatively little in what happens afterward, despite software updates, algorithm changes, and shifting patient populations silently degrading performance.
She proposed a risk-stratified framework covering clinical evidence review, acceptance testing, optional pilots, and continuous post-deployment monitoring across three domains: technical (uptime, latency), clinical (output drift, concordance rates), and impact (workflow efficiency, turnaround times).
"If we want to ensure long-term safety of AI in the world where the only constant is change, then I think we do need post-deployment monitoring to be able to do that.”
The Human Cost of AI Implementation
Dr. Susan Cheng Shelmerdine, a radiologist with subspecialty expertise in pediatric imaging at Great Ormond Street Hospital, London, shifted the focus to human factors. She presented a successful chest X-ray AI triage rollout at a district general hospital in southwest London that significantly cut cancer pathway diagnostics down from six days to three and a half days. Yet, despite positive clinical metrics, the tool had to be withdrawn in 2025 due to a lack of long-term funding.
Radiographers found themselves unexpectedly responsible for informing patients of potentially abnormal findings in real time, while AI alerts disrupted reporters’ mid-session. When the tool was removed with only one week's notice after two years of use.
"I feel that maybe when the AI was taken away, it caused some sense of moral injury to the staff in the department because they had worked in good faith to try to improve patient care," she said. Shelmerdine called for responsible exit strategies, including skills maintenance and graduated step-downs to be planned, stating that withdrawal must be managed as carefully as adoption.
The Economics Problem
The closing panel discussion covered a challenge that underlies all three presentations: proving the financial value of AI. Harvey noted that his own research and parallel work from Van Leeuwen's group had reached the same conclusion: "There's not enough evidence that AI is driving a significant cost saving for health care systems." Since AI does not currently reduce staffing, the financial case remains difficult to make. When it comes to liability, it still clearly lies with the radiologist since all of these devices require a human sign off on their output.
