Navigating the Future: Challenges and Promises of AI in Radiology

Artificial intelligence (AI) has garnered considerable attention in radiology, and the potential for 'faster, cheaper, safer' makes it a dynamic and exciting field. This is particularly true when it comes to reporting. Tools to aid diagnostic decision-making are in the marketplace, with CE marks and legislative approval. Given this landscape, the possibility of independent diagnostic decision-making AI does not seem far away, which is attractive in environments with limited resources and perhaps a more relaxed legislative environment. Diagnostic accuracy and workflow efficiency improvements ultimately promise to deliver better patient care. However, the introduction is not without its challenges, and radiology grapples with complexities as it navigates the path toward a more AI-inclusive future.

The key questions when procuring such tools are familiar, although nuanced in an AI context. Does it work in a population I plan to use it with, and how will you demonstrate that it works? It is unknown if a mammography tool trained using central European breast will work on sub-Saharan African breasts. Moreover, how might a vendor demonstrate that it will work in any given population? It may be that differences between populations in the context of the AI tool are minimal, and the tool's performance does not vary between populations. Still, thus far, this is generally unknown. The next challenge is how the tool might fit into the existing workflows and how the AI's 'view' is presented to the current practitioners. Without onboarding these critical stakeholders, gaining traction for any tool would be difficult. Various techniques, such as heat maps or ROIs, are available to highlight areas within an image for special attention to simple yes-no responses. Radiologists are accustomed to a well-established image presentation and interpretation routine, relying on their expertise and experience. Performance gains may differ depending on how and when information is presented. Adapting to these new tools demands a careful balance between technological innovation and the preservation of familiar diagnostic methodologies, at least in the first instance. Seldomly considered is performance over time. AI tools will not exist in a fixed ecosystem, particularly in imaging, over time. Acquisition systems will be replaced, software updates will be performed, and acquisition protocols will be changed. Maintaining performance over time will be an essential operational characteristic if successful implementation is to take place.

If these challenges are to be met, substantial amounts of high-quality data are needed to train and test AI algorithms. AI models in radiology heavily depend on large datasets to learn and generalise from diverse cases. However, curating such datasets is a time-consuming and resource-intensive task.

The availability of annotated images, covering a broad spectrum of pathologies and variations, is crucial for accurately training AI models. The challenge lies in collecting these datasets and ensuring they represent the diverse patient populations encountered in real-world clinical settings. An alternative model would be training and testing models in restricted populations but only badge the tool for that population, producing new models for each population. Given the difficulty of curating such datasets and the relatively easy access to AI training software and methodology, the value of the data is heightened, confirming Clive Humby’s mantra that “data is the new oil.” And, like oil, it isn’t useful in its raw state. It needs to be refined, processed and turned into something beneficial. Data’s value lies in its potential. In the coming years, healthcare systems that can curate and provide access to such data and link that data to outcomes will be in advantageous positions—healthcare data superpowers.

The challenge of getting the stakeholders on board to effect change has multiple dimensions. Advanced diagnostic practitioners in radiology have invested in their training and gained experience over many years. AI may be seen as putting them out of a Job or supercharging their productivity. The view taken will depend on the health system’s structure and incentives. These highly trained individuals are naturally curious, and the question of ‘how does it work?’ is common and generic across all AI fields. Radiologists want to understand the rationale behind a diagnosis conclusion, and trust in AI tools hinges on comprehending the decision-making process. Some AI models, especially deep learning algorithms, are complex and are often considered “black boxes.” Explaining how an AI system arrived at a specific diagnosis is important for gaining radiologists’ trust and acceptance. However, this may not be possible or appropriate, with IP being intrinsic in that explanation.

Ethics and AI have been the subject of intergovernmental summits, with attempts to establish a broad framework. Patient privacy and security are paramount concerns, especially when dealing with sensitive information. Researchers have found access to data difficult. The data guardians have built a robust framework for data governance and compliance with privacy regulations. This framework is not without its flaws. The right balance between realising the benefits of AI and safeguarding confidentiality requires a collaborative effort between data providers, patient groups, developers, and regulatory bodies. The increasing use of data safe havens for development and testing may go some way to improve the situation. Data safe havens allow those responsible for the data to control access security and standards while allowing researchers and developers access, essentially allowing the use of the data without unfettered access; they don’t get to take the data home with anything being removed from the haven controlled by the guardians.

Algorithm bias is another significant ethical challenge in AI integration. The argument is that AI algorithms trained on biased datasets may perpetuate healthcare disparities. For instance, if a training dataset does not represent demographics, the AI model may not perform equally well across groups. Inclusivity and fairness in AI development is a noble aim to ensure that these tools benefit all patients, regardless of their demographic. However, there is a requirement for a critical mass of data in any given group. It is unlikely that all AI tools will be ‘all things to all people’ shortly.

Despite these challenges, AI’s probable benefits in radiology are immense. AI has the potential for multiple gains in diagnostic accuracy, reduced workload, and faster workflows. Collaboration between industry, healthcare professionals, and policymakers is vital to navigating these challenges and ensuring responsible and effective AI integration.