By Prof. Erik Ranschaert, visiting professor at Ghent University and board member of EuSoMII – European Society of Medical Imaging Informatics
The landscape of radiology is undergoing a profound transformation. As someone who has been deeply involved in both clinical practice and the advancement of artificial intelligence in medical imaging, I’ve observed firsthand how Computer-Aided Detection (CAD) systems have evolved from simple assistance tools to increasingly sophisticated solutions. However, we’re now standing at the threshold of an even more significant change: the emergence of autonomous AI systems in radiology.
The current state of CAD and its limitations
While CAD systems have undoubtedly made valuable contributions to radiological practice, they primarily serve as supportive tools, helping radiologists identify potential abnormalities or providing second opinions. Despite substantial investments—reportedly $5.6 billion[1] in recent years—these systems haven’t fully addressed the fundamental challenges facing our field: increasing workloads, growing backlogs, and persistent workforce shortages. The shortage of radiologists to interpret chest X-rays, for example, means that many go unreported.
The reality is that traditional CAD solutions, while helpful, often provide relatively modest time savings and do little to address the major operational challenges that radiologists and healthcare institutions face daily. We need to think beyond mere assistance and consider solutions that can fundamentally transform our workflow.
The promise of autonomous AI
Autonomous AI represents the next logical step in this evolution. Unlike traditional CAD systems, autonomous AI can function independently for specific tasks, making clinical decisions without direct human intervention. This is particularly relevant for chest X-rays, which represent the second-largest imaging volume globally, with over 875 million scans completed in 2022[2] alone.
At present, Oxipit stands as the only company with CE-approved autonomous AI solution for chest X-rays, marking a significant milestone in the field. This achievement demonstrates that autonomous AI is not just a theoretical concept but a practical reality that’s already being used globally.
The practical impact on workflow
From my experience, working with healthcare institutions across Europe, I’ve observed that autonomous AI can safely filter out between 20% and 40% of chest X-rays in a general hospital setting—sometimes even more. These are cases that show no clinically significant findings and would otherwise require valuable radiologist time to review and report.
This capability creates a compelling business case for healthcare providers, particularly in environments where resources are stretched thin. The system’s ability to automatically handle normal cases allows radiologists to shift their focus and expertise to more complex cases that truly require human interpretation and deliver quicker patient care.
Addressing implementation challenges
However, implementing autonomous AI isn’t without its challenges and we need to address them to have wider adoption for this new technology. Based on my interactions with healthcare providers across Europe, I’ve identified several key considerations:
1. Regional variations: The acceptance and implementation of autonomous AI can vary significantly between countries. We’ll need to get regional regulations sorted if we want this technology to take off. With every country having their own rules or no rules at all, it’s tough to get everyone on board. It could be that we set up some best practices in a few countries first, and then other places can follow their lead.
2. Ethical considerations: As AI systems take on more decision-making roles, ethical considerations surrounding bias, fairness and accountability become increasingly important. We need to carefully consider questions of liability, responsibility, and patient consent in each context.
3. Building trust: Creating confidence in autonomous systems requires robust evidence and validation. This includes extensive fine-tuning at implementation sites to ensure false negatives are minimal or within acceptable ranges.
4. Workflow integration: The technology must be seamlessly implemented in a way that enhances rather than disrupts existing workflows. This might include different approaches for different settings—seamless integration directly into Picture Archiving and Communication System (PACS), or via an AI marketplace that serves as a single source for all applications that healthcare institutions use. It is also crucial to inform, educate, and train the recipients of the reports. There is a critical question of ensuring their preparedness, awareness, and acceptance of the technology.
5. Evaluation of results: Continuous monitoring and evaluation of AI system performance is essential to maintain accuracy and prevent any degradation over time. AI models can experience drift or decay in their performance due to various factors, such as changes in the data distribution, updates to the underlying software or hardware, or shifts in the real-world environment. Without ongoing monitoring, these performance issues may go unnoticed, leading to potential errors or inaccuracies in the AI system’s outputs. To avoid such shifts or decay, healthcare organisations should implement robust monitoring and evaluation processes.
The future of radiological practice
Autonomous AI is facing the same challenges that today’s global AI applications once had. Looking ahead, I believe we’re moving toward a future where autonomous AI will handle an increasing proportion of routine cases, allowing radiologists to focus on more complex examinations and consultative roles. This shift won’t make radiologists obsolete; rather, it will enable us to practice at the top of our expertise.
The integration of foundation models and their ability to learn from vast amounts of unstructured data—including images, reports, and clinical information—suggests we’re moving toward even more sophisticated systems. These could eventually generate comprehensive reports that incorporate all available patient information, supporting truly personalised, precision care.
The path forward
To successfully integrate autonomous AI into radiological practice, we need to generate robust evidence by continuing to conduct both retrospective and prospective studies to demonstrate the technology’s effectiveness and safety for application of these tools in clinical practice. We must adapt training to ensure that radiologists maintain their core diagnostic skills while learning to work effectively with autonomous systems. It is essential to develop clear protocols by establishing guidelines that can be adapted to various healthcare systems and regulatory frameworks. These protocols should address how to handle high-impact false positive findings and specify who will be responsible for validating and signing off on the reports. Above all, we must focus on quality by maintaining rigorous assurance processes to ensure that autonomous systems perform at or above human-level accuracy.
The transition from CAD to autonomous AI represents more than just a technological advancement—it’s a fundamental shift in how we approach radiological practice. While CAD systems have paved the way, autonomous AI offers the potential to address the core challenges facing our field: increasing demand, resource constraints, and the need for more efficient workflows.
As the only CE-approved autonomous AI solution currently available for chest X-rays, Oxipit’s implementation success demonstrates that this technology is not just theoretical but practical and beneficial. Its ChestLink solution will only produce automated reports for chest X-rays where it is highly confident that the images feature no abnormalities. This allows the sensitivity metric to be higher than 99%, ensuring high confidence in ruling out abnormalities before generating autonomous reports. This AI system which is capable of classifying images as normal or abnormal, will be a valuable improvement in addressing the chronic shortage of radiologists. However, we must approach this transition thoughtfully, ensuring that we maintain high standards of patient care while adapting our practices to leverage these new capabilities.
The future of radiology lies not in replacing human expertise but in augmenting it with autonomous systems that can handle routine cases, allowing radiologists to focus on complex cases and consultative roles. This evolution will require continued collaboration between clinicians, technology providers and healthcare institutions, but the potential benefits—improved efficiency, reduced backlogs, and better patient care—make this journey worthwhile.
[1] AI in Medical Imaging Company and Product Database – 2024. Signify Research
[2] Diagnostic Imaging Procedure Volumes Database – 2024, Signify Research