Automation Bias in Radiology: An Interview with Professor Erik Ranschaert

In the fast-evolving landscape of radiology, the integration of artificial intelligence (AI) has revolutionised diagnostic practices, offering unprecedented opportunities for efficiency and accuracy. However, with these advancements comes a critical consideration: the potential for automation bias. As we delve into the intersection of AI and radiology, we will explore this phenomenon with an expert in the field. With a wealth of experience spanning healthcare, imaging informatics, and pioneering contributions to European teleradiology, Professor Erik Ranschaert, Radiology AI Solutions adviser at Unilabs AI Centre of Excellence, is uniquely positioned to shed light on the complexities of automation bias in radiology. 

How has the integration of AI impacted radiologist decision-making based on your experience? 

Indeed, the experience or ‘acceptance’ of AI solutions can vary greatly within the group of users. The attitude towards this type of solution will also determine how they deal with the outcome. Academic research has also already shown that most of the users risk falling into the trap of the so-called ‘automation bias’. Others may, for example, systematically challenge the analyses, or disable the algorithm because they do not find this ‘intervention’ or information desirable or consider it a waste of time.  I have always been very careful, considering this an essential part of the implementation strategy. It is crucial during the implementation process to communicate well among colleagues and with other staff inside and outside the department, about what kind of solution you are going to implement and what its capabilities and limitations are. It is very important to facilitate feedback and regularly evaluate how the application affects each user's workflow. This also requires setting up a structure that allows regular consultation. Based on that feedback, adjustments or improvements can be made and this information is also important for suppliers. From the vendor's perspective, it is also important to organise user meetings so that they can learn from the users' experiences.  

Can you explain how automation bias affects radiology and its potential consequences for patient care? 

As I mentioned, this is a topic that has already been analysed with scientific research. Some articles have been published recently, including several studies on the impact AI can have on the analysis of mammography examinations and the use of the BI-RADS scoring system. The results show that radiologists reading mammograms are prone to automation bias when supported by an AI-based system, regardless of their experience level. This automation bias can be accentuated when a radiologist is fatigued or when there is a limited radiological workforce and thus limited capacity to oversee AI output. Patient safety risks will also be higher when autonomous AI systems are implemented or when the AI system continues to learn and adapt over time. In these situations, the need to assess and monitor the performance of AI systems becomes proportionally greater. These and other effects of human-machine interaction must be considered to ensure safe deployment and accurate diagnostic performance when combining human readers and AI. Recently a multi-society statement, including the ESR, has been published about this in several major journals. 

How can radiologists minimise automation bias while incorporating AI into their diagnostic processes? 

Minimising automation bias when integrating AI into diagnostic processes is crucial to ensure radiologists retain their clinical judgement and decision-making skills. There are several strategies to achieve this: 

Education and training are a first option. Radiologists should be educated on the capabilities and limitations of the AI algorithms they use. Understanding how the AI system works and its potential biases can help radiologists interpret output more critically. 

AI systems should also be sufficiently transparent. Radiologists should have access to information about how the AI arrived at its conclusions, including the features used and the confidence level of the predictions. 

Radiologists should verify AI-generated findings by comparing them with their own observations and interpretations. It is also important to conduct validation studies to assess the accuracy and reliability of AI algorithms in real clinical settings. This should be done not only locally, but also on a larger scale in multiple centres and even internationally, preferably prospectively, so that one can assess how these algorithms function in a real-world environment. 

Integration of clinical context remains important. Radiologists should integrate AI findings into the broader clinical context of each case, considering the patient's history, symptoms, and other relevant factors. AI should be seen as a tool to augment rather than replace clinical expertise. 

Finally, it is also important to obtain continuous feedback, as I mentioned earlier. Radiologists should provide feedback to AI developers based on their experiences of using the technology. This feedback loop can help developers improve AI algorithms and address problems or biases that arise during clinical use. 

One could argue that radiologists should maintain a healthy scepticism towards AI results and avoid blindly accepting algorithmic recommendations without critical evaluation. They should be aware of possible biases in AI algorithms and consider alternative interpretations. Collaboration can help reduce individual biases and promote more accurate diagnostic decision-making. 

By implementing these strategies, radiologists can effectively minimise automation bias while integrating AI into their diagnostic processes, allowing them to continue to provide high-quality care to patients based on their clinical expertise and judgement. 

What future developments in AI do you anticipate could help address automation bias and enhance diagnostic accuracy in radiology? 

Several future developments in AI have the potential to address automation biases and improve diagnostic accuracy in radiology.  

So-called "Explainable AI" (XAI) will enable AI systems to provide transparent explanations for their decisions. Radiologists will gain a better understanding of how AI algorithms arrive at their conclusions, reducing automation bias and increasing trust in AI-generated findings. 

Advanced AI models capable of estimating uncertainty in their predictions will help radiologists estimate the reliability of AI-generated findings. Radiologists can use uncertainty estimates to prioritise cases for further assessment and validation, reducing the risk of automation bias. 

AI systems that continuously adapt and learn from the feedback provided by radiologists will improve over time and better fit clinical practice. These adaptive learning systems can dynamically adjust their algorithms to address biases and improve diagnostic accuracy. 

Integration of AI with multiple imaging modalities and other clinical data sources will provide a more comprehensive picture of patient health. By analysing diverse data sets, AI systems can generate more accurate and contextually relevant diagnostic insights, reducing the likelihood of automation bias. 

AI-driven approaches to personalised medicine will enable tailored diagnostic and treatment strategies based on individual patient characteristics. By considering patient-specific factors, such as genetics, lifestyle, and medical history, AI systems can generate more accurate and personalised diagnostic recommendations, reducing the influence of bias. 

The implementation of robust ethical AI frameworks and guidelines specific to radiology will ensure responsible and equitable deployment of AI technologies. Ethical considerations, such as limiting bias, fairness, privacy, and accountability, will be integrated into the development and use of AI systems in radiology practice.