Radiologists and healthcare administrators have been waiting a long time for the arrival of promised AI tools for radiology. Most AI applications used by radiology providers today are focused on worklist prioritization. While improving the speed of patient care is incredibly important, recent strides in quality improvement open a new world of possibility.
Half a decade after being told that AI will replace radiologists, we’re instead seeing AI become a backstop for human error.
The pace of AI advancements in quality assurance (QA) is accelerating. New models are poised to make significant in-roads in 2021 — exponentially increasing the rate at which AI becomes embedded in the daily workflow of radiologists.
vRad’s AI development team recently began piloting an AI-enabled QA process with three AI models reviewing all chest CT images arriving in our system for pulmonary embolism, aortic dissection, and pneumoperitoneum — looking for discrepancies between the radiologist report and the images. Cases with possible misses are sent to vRad’s clinical QA committee for appropriate follow up and the initial results have been very promising with several critical pathologies identified by AI and confirmed by the committee.
The success of these new models nudges the door open to a flood of exciting questions:
- Can AI quality checks become near-real-time, warning the radiologist of a possible miss?
- Can we significantly reduce misses by identifying them early in the process?
- Can we keep more patients in-network and improve their care by identifying studies with difficult-to-find conditions?
- Can we expand to other AI models and conditions in the never-ending quest of improving patient care?
The answer to each of these is getting closer to yes.
The real power to improve radiology quality comes from AI’s innate ability to perform at massive scale. Instead of a practice overreading 1% of studies, AI can overread 100% of studies, looking for specific misses. Our quality improvement pilot has started in the shape of overreads for the QA team, but our sights are set on exponentially augmenting radiologist performance during the read — when the opportunity to improve clinical quality and patient care is greatest. Additionally, the immediate feedback becomes a powerful tool for radiologists to continually sharpen their skills, especially those in the earlier stages of their career.
Quality based AI models will provide concrete metrics, such as number of misses per practice or number of patient lives saved by quick feedback. These quantifications will be leveraged for quality improvement across the clinical aspect of radiology as well as the underlying business of radiology. For example, a near-miss pulmonary embolism or hip fracture that is caught by AI will not only ensure better patient care, but also that the patient does not need to seek care elsewhere — ultimately driving down patient expenses and lost revenue for the hospital.
As practices learn the nuances of AI and deepen their understanding of the tools, AI will provide additional in-read enhancements like immediately measuring nodules, making automatic statements in reports, providing visual indicators including “red dot” notifications or providing visualization masks over probable pathology.
I believe that utilizing AI to improve the quality of care available to patients will propel even further advancements in the technology and drive more widespread adoption quickly. I’m excited that the vision of an AI augmented reading room continues to unfold.
