Computer Vision in Radiology: How ML Will Disrupt the Future of Clinical Radiology

Working with computer vision in radiology yields potential benefits both for patients and health centers. However, some fear that machine learning (ML) can outperform physicians and may eventually replace them. 

The main driving force behind the emergence of AI in medical imaging has been the pursuit of greater efficiency and effectiveness in clinical care. Recent advances in technology, especially in ML, are going to change radiology as we know it. And it’s not limited to image interpretation. ML will shape the entire data pipeline — from ordering an X-Ray or MRI scan, to making a diagnosis. Radiological imaging data continues to grow faster than the number of trained physicians. With the active introduction of algorithms and models into the field of radiology, a logical question arises — will AI replace radiologists?

Will ML replace radiologists?

In 2016, Ezekiel Emanuel, a renowned American oncologist, made a bold statement that machine learning will replace much of the work of radiologists and pathologists in the coming years. His words were picked up by authoritative medical publications, causing concern among recent medical school graduates. Even though many of these experts have already changed their minds about machine learning for radiology, some medical students still choose not to specialize in radiology because they are concerned that this occupation will be taken over by AI. In a study carried out in 2018, although only 29.3% of Canadian medical students agreed that AI may replace radiologists in the future, 67.7% believed that AI would reduce the demand for radiologists. 

Indeed, tech advances show that AI will play an important role in radiology. Overall, artificial intelligence in the healthcare market is projected to grow from $6.9bn in 2021 to $67.4bn by 2027 — a compound annual growth rate of 46.2% — and medical imaging will benefit from technology that can read and analyze many images in a short time. 

Does this mean that the radiologist as a profession will disappear forever within the next five to 10 years? It’s unlikely to happen. And here’s why.

Вeing a radiologist is more than image reading and interpretation

Radiologists do much more than just interpret the images they receive. Specialists train machines to recognize abnormal tissues on a CT chest scan or hemorrhages on an MRI of the brain. However, AI is still not capable of fully performing all the tasks of a radiologist. The job of interpreting images is just one component. Radiologists constantly consult other doctors, provide ablation therapy, etc. Don’t forget about the human factor and empathy, as radiologists also discuss procedures and results with patients.

Machine learning for radiology is far from ready

No matter how optimistic the forecasts for the introduction of new technologies in healthcare are, the clinical processes of using images based on artificial intelligence are far from being ready for everyday use. 

Today, AI specialists train machines to help radiologists focus on multiple tasks at once. This approach will expand the scope of AI applications and accelerate the introduction of machines into current clinical practice. Of course, it will take many years to create a comprehensive collection of subjects, further expanding the role of radiologists in the world of artificial intelligence.

Inability to collect enough images

The more the machine examines cases associated with the images of patients who have been diagnosed with cancer, bone fractures, or other pathologies, the faster it learns to spot potential diseases. Unfortunately, radiology images are not held in one global repository. Their owners are patients and hospitals. It will take a few years to collect images and train models.

Regulatory changes

Last but not least, the integration of AI into medicine requires major changes in medical regulation and insurance. It is necessary to establish who, in the event of a mistake, will be held liable for a misdiagnosis — the doctor, the hospital, the AI provider, or the engineer who created the algorithm? So it is likely that AI machines will be great partners for radiologists in the future. But for this to happen, an important step needs to be taken to change regulations.

The underlying challenges of object recognition 

The greatest advantage that a machine has over humans is that it can process data endlessly and doesn’t need rest. The system can be trained to rely on thousands of different approaches for thousands of cases. However, upon occasion, it will come across a person with an extra or congenitally absent bone or hypoplastic structures and will be unable to diagnose the patient. 

This problem becomes most apparent when considering the challenges faced by the radiological investigator collecting and classifying types of anatomical structures and abnormalities found on chest X-rays. This researcher will need to obtain images and related data for a computer to demonstrate abnormalities of the heart, mediastinum, lungs, bones, pleura, and various other structures.

For a machine learning system to replicate the work of a radiologist, it would have to include a large set of specific ML algorithms, each designed to answer a specific clinical question. 

However, integrating and aligning such a wide and varied set of learning algorithms — perhaps from several different providers — into a single clinical system is likely to require a significant investment of time and effort in validation and testing, not to mention potential regulatory challenges.

Pathway for incorporating ML into radiology today

Machine learning is evolving rapidly and has the potential to be a much better technology than analyzing images by physicians. But the best systems currently available roughly correspond to human capabilities and are used only for research purposes. 

Even if the use of machine learning technology throughout society continues to grow exponentially, it is far from clear that ML algorithms in medical imaging will necessarily experience such astronomical growth. Advances in computational speed can only guarantee that the same answer — including the wrong one — can be provided 1000 times faster.

Currently, machine learning for radiology requires many well-annotated studies of images to be submitted by human researchers. The researchers then periodically test each algorithm for reliability and accuracy. Large imaging datasets will need to be collected and disseminated to institutions and radiology clinics. This is an activity that requires work and trust to overcome technological, institutional, and regulatory barriers.

According to the results of an IBM study, ML models can reach high accuracy and can be on par with medical professionals already. The question of whether diagnoses based solely on machine learning are acceptable will become an ethical issue. If this is viewed solely as a technological update, and if society has already adopted other innovations such as self-driving cars, then this change may not be controversial. On the other hand, if there is a significant negative public reaction to the loss of human interaction in the medical field, then it is possible that radiologists will not be replaced for a very long time, if at all.

The benefits of computer vision in radiology

In the field of radiology, trained physicians visually evaluate medical images and report the results to detect, characterize, and monitor diseases. This assessment is often based on education and experience and can sometimes be subjective. In contrast to such qualitative reasoning, AI is best at recognizing complex patterns in image data and can automatically quantify them. More accurate and reproducible radiological assessments can be made when AI is integrated into the clinical workflow as a tool to assist clinicians.

For example, skin cancer can be difficult to detect early because the symptoms often mimic those of common skin conditions. With the rapid advancement of technology in the near future, medical computer vision systems will be used to effectively distinguish between cancerous skin lesions and benign lesions. Trained with an extensive database of images of both healthy and cancerous tissue, it can help automate the identification process and reduce the likelihood of human error.

While the patient is undergoing a CT angiography or an MRI scan performed by a machine, the radiologist will be able to consult with another doctor and other members of the clinical team. This will significantly speed up the working process and increase the effectiveness of the treatment.

Machine learning for radiology will give physicians more time to think about what is going on with patients, diagnose more complex cases, collaborate with patient care teams, and perform invasive procedures.

Final thoughts

The integration of artificial intelligence with the knowledge and skills of radiologists will bring significant medical benefits and increased productivity. While machines perform routine tasks, radiologists will spend more time consulting other doctors about diagnoses and treatment strategies. 

If the predicted innovations in deep learning image analysis are implemented, service providers, patients, and payers will gravitate toward doctors who have figured out how to work effectively with ML models in radiology.

Postindustria offers a full range of machine learning services and solutions. Please leave us your contact details and we’ll get in touch to discuss your custom solution.

Learn how AI transforms healthcare

Download the eBook