Developing AI to Optimize Radiology

Advancing technology has always been a crucial part of medical innovation and radiology is no exception. Despite breakthroughs in imaging technology, radiology is a field increasingly burdened with heavy clinical, research, and administrative demands. Just on the clinical side, it’s estimated that some 95 percent of patients who enter a hospital need imaging, and the demand continues to grow even as fewer young physicians are entering the field. Radiologists require intuitive solutions that will enable them to treat more patients with greater diagnostic certainty while removing process friction and non-value added rote activities—that’s where artificial intelligence (AI) comes in.

Building the case for AI in radiology

Although there are a few skeptics of AI, clinicians are increasingly realizing that it could be a key tool in managing the demands of the modern-day radiologist. Paul Chang, MD, a radiologist and informatics expert at the University of the Chicago School of Medicine, is a frequent speaker on the topic of imaging AI. At the annual 2017 Healthcare Information and Management Systems Society (HIMSS) meeting, in a session titled "Leveraging IT to Optimize Quality in Radiology,” Dr. Chang said there is no reason for radiologists to be apprehensive of AI. The technology, he assured, will allow radiologists to spend more time doing the things they like (i.e., referring physician consultations) and less time doing things they don’t like (i.e., reviewing 100 chest x-rays) and those tasks are typically ones where it’s easier to make a mistake. He insisted that the underlying deep learning that drives AI will enhance and reinvent the practice of radiology and that should be celebrated.

For example, in a prostate cancer case study survey conducted by Sectra in three key markets, radiologists agreed that AI application would positively benefit daily workflow. The survey found that 68 percent of radiologists agreed that, “automatic characterization (AI) and scoring lesions according to internationally accepted criteria would be valuable to me.”

Moreover, a recent commentary in Academic Radiology titled “Toward Augmented Radiologists” supports an optimistic view of imaging AI. The radiologist authors at Massachusetts General Hospital acknowledge there will be a short-term learning curve for radiologists – both residents and their supervising radiologists. But in the long-term, “artificially intelligent software assistants” will transform the teaching of radiology residents and offer new opportunities for attending radiologists in their practice as well. The bottom line, according to the authors: AI will serve radiologists well.

Creating comprehensive AI solutions

The potential for AI solutions is becoming more and more apparent and real. GE Healthcare is one company that is committed to developing AI tools to assist not only radiologists but also departmental users in their workflow. Earlier in 2018, the company sold a portion of its healthcare IT business for just over $1 billion to help increase investment in AI, smart diagnostics, and enterprise imaging. GE Healthcare's focus on this sector is part of a wider trend. It's projected that AI healthcare applications will grow from $216 million in 2016 to $10 billion by 2026.

The company is also partnering with academic medical institutions in more comprehensive ways to integrate AI into clinical practice – and imaging is just part of the focus. GE Healthcare and Partners HealthCare signed an agreement to “integrate (AI) into every aspect of the patient journey.” Chief Data Science Officer Keith Dreyer, DO, a radiologist, mathematician and computer science expert, said the goal is to better put data to work for patients and physicians. According to Dr. Dreyer, a typical hospital generates enough data per year to fill 20 million four-drawer filing cabinets, but 97 percent of that data never gets used. His take is that AI is essential as there is just too much data in healthcare for people to analyze.     

Dr. Dreyer cited AI application for stroke diagnosis. If, for example, a hospital performs 200,000 MRI exams a year with 10 percent of the scan interpreting a stroke, the entire database can be used to build a diagnostic algorithm. Such an application could then be deployed to quickly alert radiologists of a suspected stroke and significantly improve door-to-needle time.

These recent advances show how AI benefits radiologists now and that future advances will enable them to deliver better patient care and continue to drive advances in their field.