How will AI revolutionize medical imaging?

future health index
Dr. Declan O'Regan

Sep 19, 2017 - reading time 5 mins

By Dr. Declan O'Regan

Reader in Imaging Sciences, MRC London Institute of Medical Sciences


Dr Declan O’Regan is a Consultant Cardiac Radiologist and Reader in Imaging Sciences at the MRC London Institute of Medical Sciences within Imperial College London. He leads the Institute’s Robert Steiner MRI Facility and is also Director of Imaging Research at Imperial College Healthcare NHS Trust. His research is focussed on applying machine learning approaches to clinical imaging for understanding the genetic basis of heart disease and predicting patient outcomes. His multidisciplinary research is supported by the British Heart Foundation, Medical Research Council, and National Institute for Health research. He has a number of research collaborations worldwide and has recently been awarded the Roentgen (RCR) and Rohan-Williams (RANZCR) travelling professorships for a series of international talks on the use of artificial intelligence in radiology.

Dr. Declan O'ReganClick here to read less

fhi logo mobile

So-called visionaries imagine a world where artificial intelligence (AI) replaces many of the roles that clinicians currently fulfil, but their predictions often amount to little more than banal futurism. AI is set to radically change the interaction between doctors and their patients, however, its impact will be driven by economics as much as technology and will be progressive rather than revolutionary.

Doctor and nurse reviewing medical record in hospital

Tech investors have focused on radiology as a domain ripe for developing specialist algorithms that can interpret images. This is unquestionably one of the most exciting areas of biomedical research but the initial advances brought by AI will be less glamorous and directed at ensuring we get the most value out of the resources we already have.

Here are my five tentative predictions for how technology will change the way diagnostic imaging is conceived and delivered:

1. Management of human resources to maximize use of scanning capacity

The UK’s National Health Service (NHS) handles over 40 million imaging investigations per year and yet the technology used to manage demand and optimize efficiency is archaic compared to comparable industries. AI has the capability of developing sophisticated scheduling algorithms that dynamically match the needs of patients to the capacity of resources. After the patient has been investigated, reporting radiologists typically work in rigid linear workflows guided by a narrow interpretation of clinical urgency. The development of regional networks could provide the critical mass for AI systems to manage the distribution of image interpretation, enabling the anticipation of fluctuating demand, making dynamic responses in reporting bandwidth and optimizing prioritization of team workload.

Imaging is a rich resource of prognostically-valuable data but making accurate predictions about the future progression of a disease depends on multiple interacting features at different scales that may be beyond human perception.

2. Decision support for radiologists

At least in the medium-term, radiologists will continue to perform the vast majority of image interpretation. Aside from a trained human’s capability for interpreting complex and subtle findings on diagnostic imaging, radiologists use their clinical experience to provide meaningful actionable reports. Nonetheless, discrepancies in reporting remain frequent and human factors are attributable in the majority of cases. Short of direct image analysis, AI offers a wealth of approaches to improve safety margins such as learning to identify cases that are at higher risk of error and initiating selective double reporting. Natural language processing and use of image metadata could also enable context-sensitive prompting to avert misdiagnosis with seamless linking to educational resources. Such additional safeguards could be an innovative step towards reducing error and malpractice costs by providing a decision support framework that continuously learns from adverse events globally.

3. Risk stratification

Clinicians are poor at predicting outcomes and this can lead to ineffective management choices. Imaging is a rich resource of prognostically-valuable data but making accurate predictions about the future progression of a disease depends on multiple interacting features at different scales that may be beyond human perception. Automated risk stratification could be a powerful tool for identifying high-risk patients so that management of their care can be optimized. Supervised machine learning is an unbiased technique that searches for patterns in data that are most strongly aligned with an outcome of interest – such as expected survival or response to a specific therapy. These algorithms can be trained on any population where the endpoint is known and then applied to new patients for personalized risk-stratification. Examples of survival prediction include my group’s work on 3D analysis of cardiac motion in heart failure and others’ research on imaging characteristics of tumors. Here machine learning enhances what imaging has to offer and expands the role of radiology in prognosis far beyond routine disease staging. Perhaps in the future we will come to see radiologists as clinical informaticians that integrate multimodality health data and recommend treatment strategies.

4. Automated image interpretation

A general artificial intelligence capable of interpreting a request, performing sophisticated perceptual and classification tasks, and then formulating a meaningful natural language report is some way off. These tasks are readily performed by humans but are currently very challenging for machines. More well-defined tasks where large training datasets are available could still be transformed by AI, and early successes are already coming from first reading of mammograms and automated interpretation of abdominal CT. The challenge will be in ensuring the false positive rate doesn’t lead to inefficiency and over-diagnosis. Presumably, the common standard for negligence will be that expected of a reasonably competent doctor, but may evolve for both humans and machines to that of the best performing algorithm.

5. Global data sharing

Machine learning is only as good as the quality of the training data – and the current vogue for deep learning networks requires very large datasets to prevent over-fitting. The only route to substantial progress is the development of accessible global registries that link image data with diagnoses and outcomes. As a community of healthcare users and providers we need to accept that the benefits of well-managed data sharing vastly outweigh the potential risks if there is robust protection of confidentiality and safeguards to prevent restrictive commercial exploitation. International standards for the validation of AI in medicine must be established and be a requirement for publication. The American College of Radiology should be congratulated for establishing a Data Science Institute to guide the appropriate development and implementation of AI tools to help radiologists improve medical imaging care.


Autopilots have not replaced humans and have been instrumental in improving the precision, economy and safety of air travel. The same might be expected for the use of automation in medical imaging if appropriate safeguards are in place. Instead of facing extinction, radiologists will be pivotal in this information revolution. If the evidence base is sound we owe it to our patients to embrace new AI technology and be pioneers of its adoption in healthcare.

Share this article

  • Link copied


You are about to visit a Philips global content page


You are about to visit the Philips USA website.

I understand

Our site can best be viewed with the latest version of Microsoft Edge, Google Chrome or Firefox.