I’ve been a radiologist for 14 years, a doctor for almost 20 years and a programmer for 30 years. I’ve witnessed a lot of technological change in that time but perhaps none as controversial as, or holding the promise of, artificial intelligence (AI). Public interest in the use of AI has exploded since ChatGPT went mainstream in 2023 and over the last handful of years a plethora of AI tools designed for use in diagnostic imaging have emerged, been certified for use in the NHS and have been deployed.
Despite the promise these tools hold and the impressive feats they have already demonstrated, there is still a lot of scepticism amongst radiologists and related professionals with the propagation of several common myths. I would like to address some of these misconceptions in this post.
Myth 1: AI will replace radiologists
This is perhaps the most pervasive myth in our field. This is certainly not the case at present and evidence suggests that AI is much more likely to augment, rather than replace, radiologists. It has been shown repeatedly that whilst AI models can match radiologists in certain tasks (our own whitepaper showed a 93.5% concordance between AI and human), the highest accuracy is achieved when AI and radiologists worked in tandem. This synergy between human expertise and machine learning is likely to redefine our role rather than eliminate it. With current workforce challenges (which show no signs of abating), AI models can help us report faster, more accurately and more safely.
Myth 2: AI is a “black box” that can’t be explained
There was a time when this statement was true but great strides have been made to develop more interpretable models, in fact there is an entire field known as “explainable AI (XAI)” which exists to try to make the internal workings of neural networks more transparent. A detailed discussion about this field is beyond the scope of this article but there are various techniques including:
- Saliency maps: Which part of an image was most influential to the decision?
- Adversarial examples: These are inputs to an AI model that are intentionally designed to cause the model to make a mistake.
- Network dissection: Analysing groups of neurons (essentially mathematical functions) to understand what concepts in the model they represent.
Good governance around tool selection including validation of published vendor data (something we do with the AI tools used here at Medica) goes a long way to alleviate some of the mystique surrounding these tools.
Myth 3: Implementing AI is too complex and costly for the NHS
Whilst initial implementation can be challenging and expensive, evidence strongly suggests that AI is cost-effective in the long run. There are alternatives to deploying models on premise yourself and there are advantages to outsourcing the management of AI tools to third parties such as Medica. AI tools can increase scan and report throughput in departments, lead to better utilisation of reporting time through better scan allocation and save money through earlier detection of disease.
Myth 4: AI will make radiology reports standardised and impersonal
I disagree strongly with this concept. AI has the potential to enhance personalised care greatly. By offloading routine tasks to AI, departments can free up radiology time to focus on more complex cases and better patient interaction. Radiologists could choose to spend time saved by AI tools (e.g. with the autodetection, segmentation and volume computation of lung nodules or liver lesions) to analyse other parts of the study more closely or to spend longer assimilating a patient’s complex history via previous reports.
Myth 5: AI can work independently without human oversight
This is a dangerous misconception. Whilst AI tools have made remarkable progress recently, the need for human oversight remains critical and here’s why:
- Contextual understanding: No matter how advanced an AI model is, it lacks the contextual understanding that clinicians possess.
- Edge cases / rare conditions: AI models are trained on very large datasets but these (understandably) are full of the most common conditions. Given that the current state of the art in AI relies on massive data input, AI models frequently fail to diagnose uncommon disease states or unusual presentations.
- Quality control: Human oversight is essential for maintaining the quality & safety of AI-assisted diagnoses. “Model-drift” is a one such example of a phenomenon that can occur over time with AI models whereby the local population the AI is deployed upon can deviate over time from the training set leading to a reduction in specificity and sensitivity. This needs to be monitored for and corrected – currently a process that requires human intervention.
- Ethical considerations: Diagnostic decisions often involve complex ethical considerations that AI is simply not equipped to handle. Radiologists will continue to play a vital role in ensuring that AI-assisted diagnoses align with ethical standards and patient preferences.
To sum up, whilst AI deployment and usage in radiology is not without its challenges, many common fears are unfounded or exaggerated. As radiologists, our role is evolving to incorporate these new tools to improve our diagnostic accuracy and, ultimately, patient care.
I would love to hear from you any other myths or concerns you may have and we will try to answer them on the blog in a subsequent post.
Dr Garry Pettet, Clinical Safety & Audit Officer
