I strongly believe that the Game Changers in the field of AI and Healthcare are the ones who will Master the Art of seamless integration into current clinical and radiological workflows, or … total replacement & disruption of how we make it today !

Today I will talk about one clinical workflow which might be strongly influenced by AI, if not becoming radiologist-free:

The Traumatic Musculoskeletal X-Ray Pathway

X-ray or conventional radiography is the classic 2D black and white image performed to visualise bones, joints, abdomen and chest.

Traumatic injuries of the Musculoskeletal (MSK) system are the most common reason for admission in the Emergency Room and X-rays of the injured part of the body is usually the first exam performed after clinical evaluation. The number of X ray performed yearly in Western countries is in the range of hundreds of millions.

When the patient is admitted and examined in the emergency room by the emergency physician and/or the orthopaedic surgeon (we will call these doctors: the clinicians), an X-ray is performed and analysed immediately by these clinicians who have been trained to make minimal and sufficient analysis of 2D radiological exams but are not specialised in Medical Imaging. After looking at the X-ray, they treat and/or discharge the patient without the involvement of the Radiologist in the majority of the cases. And thus the immediate treatment decision is based on the X-ray analysis of the clinicians and not the radiologist.

This workflow works pretty well, however and according to my experience, some lesions are overlooked by clinicians during their first interpretation, and the absence of treatment of these lesions could be detrimental to the patients. I would say that these overlooked lesions represent less than 5 % of the cases.

Fortunately for these patients, these lesions are “usually” well detected by a trained radiologist.

So, when does the Radiologist intervene?

In one of my previous hospitals, radiologists were reviewing all of these MSK X-ray the next morning or Monday morning after the week-end (or on demand if the clinician wants an immediate radiologist confirmation).

When the radiologist enters the scene, 12 to 48 hours after the patient admission and X-ray exam, the patient might be recovering from his surgery or even back to home. The goal of the radiologist is to review the images and produce a formal complete report, and to look for abnormalities missed during the initial analysis.

When a significant abnormality is detected by the radiologist and the patient already went back home, usually he immediately informs the clinician, who calls back the patient and a treatment is then performed (with a delay).

So for each patient, the radiological exam is analysed by two to three different physicians at different time-points, and for some lesions the diagnosis and treatment are delayed until the radiologist detects it. The setting which I describe here is in a large and well staffed western world academic hospital, and I suspect that in many smaller hospitals around the world, Musculoskeletal X-rays are reviewed only by the clinician in the emergency room without any radiological expertise involved and without any formal radiological report issued. On the other hand, the task of reviewing these MSK X-ray by radiologists, hours and days after the patient admission and treatment, is perceived by many of them as frustrating, boring and repetitive, especially when they don’t find any lesion overlooked by their clinician colleagues and questions their added value.

For all these reasons, I see here a clinical & radiological workflow which could be disrupted by Deep Learning and AI.

Deep Learning is a subset of Artificial Intelligence which uses neural networks to recognise patterns on images and is able to classify images into categories with variable accuracy. Some projects describes diagnostic performance of AI Algorithms equivalent to a human radiologist with an area under the ROC curve of 0.994 in narrow tasks such as detection of hip fractures from frontal pelvic x-rays.

A Deep Learning algorithm which automatically analyses the MSK emergency X-rays right after their realisation and classify them as normal or abnormal, then according to the abnormality types, will assist the non-radiologist clinician who takes care of the patient, giving him an “expert” support & validation of his diagnosis and correcting him instantaneously in the emergency room without waiting for an Expert Radiologists review 12 to 48 hours later.

If an AI algorithm of this kind has a negative predictive value (NPV) of 100%, the clinician (junior or senior), the radiology technician or the nurse could be reassured and discharge the patient immediately if no other procedure or check ups are needed.

Thus, we could ultimately decrease the need for radiologists in these repetitive and delayed tasks, improve Just-in-Time Interventions at Point of Care, reduce the delay in diagnosis & treatment, reduce the medical error rates, save time & money, unburden overcrowded emergency services and ultimately contribute to a better healthcare system.

However, we should keep in mind that the task of reviewing these Traumatic MSK X-rays by Radiologists is also the opportunity for young doctors to train in medical imaging and it remains valuable for the Healthcare system in an indirect way.

I believe this is a good example of how AI could be beneficial for the Radiology Workflow and the Healthcare system. And as AI algorithms are becoming more and more accurate and are covering broader clinical problems, it is crucial to think about the right balance between Man and Machine to offer the best Service to our patients.

Stay tuned for the next chapter of a Realistic Future of Radiology Powered by AI!

Dr. Amine Korchi is a Swizerland-based radiologist and Associate Partner (Europe) at HS.