The following is a guest article by Morris Panner, President at Intelerad Medical Systems
By 2034, the U.S. could face a shortage of up to 124,000 physicians due, in part, to burnout. In the medical imaging field, specifically, this challenge isn’t exactly new – but the technology that could help remedy the problem is. The usage of this potential solution in the imaging space has been delayed due to the perception it’s a bit of a rule-breaker.
A fairly recent arrival to radiology, Generative AI has emerged as a powerful resource, showing promise in a variety of tasks, from supporting streamlined workflows to synthetic image generation. AI, in general, is no stranger to medical imaging. Like a high-achieving older sibling, predictive AI has long been a prized resource, used to support diagnostic capabilities and identify risk factors to help predict a patient’s likelihood of developing certain conditions. Its ability to improve detection rates is legendary in radiology circles. Studies have repeatedly detailed impressive statistics, such as a software program that reduced the miss rate of potentially cancerous lesions at eight health centers by more than half – from 32.4 percent to 15.5 percent.
Generative AI, on the other hand, has overcome obstacles to claim its place in the field. It has learned to listen and take direction, slowly gaining the trust of those who watched predictive AI set the bar high, as only a big brother can. Generative AI produces content based on data inputs provided by users. The models “learn” from the inputs they’re given, which can contain inaccurate information, leading programs to offer unreliable content or wrong answers. So historically, utilizing generative AI to guide clinical care has been risky. However, it is making strides in many other areas, especially when it comes to reducing the burden on radiologists and improving patient outcomes.
A New Era: Support Tailored to Specific Needs of Patients and Practitioners
Systems supported by generative AI can optimize workflows by triaging scans according to urgency, ensuring the rapid reading of images belonging to patients most in need of immediate intervention and treatment. It can also reduce the length of exams and enable radiologists to read and report on more imaging studies per shift, increasing turnaround time without sacrificing quality for speed. A survey conducted last year by the European Society of Radiology found that 23.4% of radiologists using AI-based algorithms for clinical workflow prioritization considered algorithms to be very helpful for reducing the workload of the medical staff, and 62.2% reported the algorithms as being moderately helpful. Furthermore, integrating AI with radiology information systems has been found to reduce turnaround time by 30%.
While math isn’t its forte, generative AI has demonstrated artistic abilities, proving adept at image reconstruction and enhancement, which has been useful in challenging cases. It creates high-quality images of organs based on low-resolution ultrasounds and scans, significantly expanding MRI capabilities. Even text-to-image generation is not out of the realm of possibility. It could serve as a promising future resource for the augmentation, manipulation, and creation of medical images. A paper published recently in the Journal of Medical Internet Research found that one generative AI model was able to create realistic x-ray images based on short text prompts. Its other achievements included reasonable reconstruction of missing aspects within a radiological image and developing a complete, full-body radiograph using a single image of a knee to start with. This is especially helpful in areas where higher-quality imaging is a challenge, such as in rural or remote settings, thus increasing global patient access to medical imaging.
Known for its ability to study, generative AI can even help others do their “homework” when it comes to patient support. Practitioners use it to develop effective treatment plans, personalizing them based on patient-specific distinctions, such as medical history and even genetic makeup. To assist patients in better understanding their treatment options and care instructions, AI can generate educational materials and, through tools like automated chats, can answer patients’ questions about medication, symptoms, and procedures.
Encroaching on Big Brother’s Diagnostic Territory
Though generative AI has shown it can enhance and create images, the interpretation of those images traditionally requires a higher level of expertise and experience than machine learning algorithms can provide. Ever diligent, though, generative AI is working hard to change that.
Training of generative AI models to provide automated diagnoses based on the analysis of scans has begun. The technology can be taught to detect abnormalities in medical images and then offer a diagnosis based on the type and severity of the abnormality. It can also support radiologists and enhance their diagnostic accuracy by providing supplemental insights and information, such as quickly identifying patterns in expansive datasets that an imager’s eyes may not immediately pick up on.
A recent controlled trial conducted by Swedish imaging experts produced promising findings for generative AI’s future in diagnostics. The study randomly allocated over 80,000 women to either a control group that underwent standard double-reading by radiologists or AI-supported assessment. The technology detected approximately 20% more cancers when compared to standard screening with no impact on false positives. And the positive effects on radiologists’ workload were undeniable, as AI reduced the mammography screen-reading workload by 44%. The researchers estimated radiologists can read an average of 50 breast imaging exams per hour. Assisted by generative AI, 40,000 scans were read quickly enough to amount to five months’ worth of saved time for radiologists.
Helping Generative AI Gain Intelligence is Key
The key to increased usage of generative AI in diagnostics is expanding its medical knowledge by increasing the size and accuracy of the data pool it has to draw from. Basically, it needs a library card and subscriptions to medical journals. Improving the resources from which it learns will sharpen its capabilities. The technology is dependent on the training data it is given as its only source of information to review and study. If provided with data that contains biases or limitations, it learns them and those inaccuracies could be reflected in the results produced. Collective generation approaches, such as ensemble methods and mixture models, have been found by researchers to effectively ensure stronger and more varied outputs. And even AI seems to benefit from teamwork, as numerous models collaborating together appear to result in more consistent results.
Intent on becoming an overachiever, generative AI can even be trained to teach itself in some instances. It is already being used to generate synthetic medical images that can be used to train machine learning models. This is helpful as real medical images that document very rare conditions, diseases, or certain imaging modalities are not always readily available.
Generative AI hasn’t yet caught up to predictive AI’s performance in medical imaging, but it’s well on its way. The evolution is also quite nuanced since both types of artificial intelligence possess different capabilities that allow them to provide value in their own respective ways. They’re not interchangeable, but when utilized properly and responsibly to manage the tasks to which each is best suited, generative and predictive AI are rapidly becoming indispensable resources. In all likelihood, the imaging community has only just begun to reap the unimaginable benefits generative AI is on track to provide.
About Morris Panner
Morris Panner is the President of Intelerad Medical Systems, leading the company in delivering better care through improved technology. Morris served as CEO of Ambra Health from 2011 until its acquisition by Intelerad in 2021. Morris is an active voice in the cloud and enterprise software arena, focused on the services and healthcare verticals. He is a frequent contributor to business, healthcare, and technology publications. Previously, Morris built and sold an industry-leading business-process software company, OpenAir, to NetSuite (NYSE:N).
