The great challenges in radiology are accuracy of diagnosis and speed. Yet for radiology, machine learning and AI systems are still in early stages. Last August, a National Institutes of Health (NIH)-organized workshop with the Radiological Society of North America (RSNA), the American College of Radiology (ACR) and The Academy for Radiology and Biomedical Imaging Research (The Academy) kickstarted work towards AI. Their goal was to collaborate in machine learning/AI applications for diagnostic medical imaging, identify knowledge gaps, and to roadmap research needs for academic research laboratories, funding agencies, professional societies, and industry.
The report of this roadmap was published in the past few days in Radiology, the RSNA journal. Research priorities in the report included:
- new image reconstruction methods that efficiently produce images suitable for human interpretation from source data
- automated image labeling and annotation methods, including information extraction from the imaging report, electronic phenotyping, and prospective structured image reporting
- new machine learning methods for clinical imaging data, such as tailored, pre-trained model architectures, and distributed machine learning methods
- machine learning methods that can explain the advice they provide to human users (so-called explainable artificial intelligence)
- validated methods for image de-identification and data sharing to facilitate wide availability of clinical imaging data sets.
Another aim is to reduce clinically important errors, estimated at 3 to 6 percent of image interpretations by radiologists. Diagnostic errors play a role in up to 10 percent of patient deaths, according to this report.
It is interesting that machine learning, more than AI, is mentioned in the RSNA materials, for instance in stating that “Machine learning algorithms will transform clinical imaging practice over the next decade. Yet, machine learning research is still in its early stages.” Radiology actually pioneered store-and-forward technology, to where radiology interpretation has been farmed out nationally and globally for many years. This countered a decline in US radiologists as a percentage of the physician workforce that started in the late 1990s and continues to today with some positive trends (Radiology 2015). Perhaps this distribution model postponed development of machine learning technologies. Also Healthcare Dive, RSNA press release