June 01, 2022

Non-interpretive Artificial Intelligence Tools in Radiology

RecommendThanks for recommending this page!

The page you recommended will be added to the "what others are reading" feed on "My ACR".

 BookmarkThanks for bookmarking this page!

The page you bookmarked will be added to the "my reading list" feed on "My ACR".

Written by Sohum Patel, BA, MS4, Geisel School of Medicine at Dartmouth
Edited by Gregg Khodorov, MD, MBA, and Ronak Ahir, BS

Sohum Patel, BA, MS4
Sohum Patel, BA, MS4
There is no denying that, in recent years, artificial intelligence (AI) in healthcare — and particularly in radiology — has been a significant source of growing excitement as well as understandable trepidation. An article from The Verge reported last month that an autonomous AI tool that can read chest radiographs without radiologist oversight passed EU regulatory clearance1; reactions ranged from celebration to disdain. However, while most public conversations have revolved around AI solutions that can recognize, diagnose or classify pathology, less attention has been paid to the use of AI tools to help solve non-interpretive problems in radiology2. From optimizing image acquisition and quality to improving efficiency in radiology practices, non-interpretive AI solutions can be leveraged in a variety of ways.

Image Acquisition and Quality

Increasing demand for imaging has raised questions regarding the tradeoff between image quality and efficiency. Conventional wisdom dictates that image quality is a function of time and/or dosage of radiation or contrast (i.e., improving quality of MR images requires longer scan time and higher radiation doses yield higher quality CT imaging). However, a variety of solutions have been developed in recent years to optimize image acquisition and reconstruction. Deep learning methods using convolutional neural networks have been shown to improve signal to noise ratio in MRI3 and allow for faster scan time without sacrificing image quality. Other deep learning tools have been used to reconstruct high-quality images from low-dose CT4 or PET5 data, reducing the need for high radiation doses to construct high-quality imaging.

Workflow and Triage

Protocoling radiology studies is key to ensuring that patients undergo the most appropriate exam given a variety of parameters, but it is also among the more tedious tasks during regular clinical practice. Manual protocoling is susceptible to human error and can be a source of inefficiency in practice. This could be another sphere in which AI solutions could optimize efficiency and outcomes. Studies have shown that natural language processing models can be used to predict and automate appropriate protocols for MRI brain6 and MSK7 exams.

Triage is another sphere in which AI could prove incredibly valuable. Recent entries to the market analyze studies for critical findings (pulmonary embolus, intracranial hemorrhage or large vessel occlusion, cervical spine fracture) then report them to radiologists so those studies can be prioritized in a worklist. However, non-interpretive worklist management might also have potential. For example, one department investigated an algorithm that calculated interpretation time based on type of study and subsequently assigned custom worklists to each radiologist based on their speed of interpretation8. They found that this method resulted in a decrease in overall interpretation time.

Implications for Students and Trainees

As a field that is fundamentally tied to technological innovation, radiology is uniquely poised to embrace a variety of innovative techniques that can improve the efficiency and quality of care delivered. AI appears to be the next frontier for that evolution. The amount of diagnostic imaging being acquired is at an all-time high, and radiologists need to be more efficient than ever before in finding life-threatening pathology without sacrificing quality. It is imperative that radiology education and training begin to incorporate AI into curricula.

While diagnostic tools have driven much of the hype, non-interpretive tools also require recognition and could represent a larger part of the AI revolution than people realize. Dr. Ryan Lee, Chair of Radiology at Einstein Network Pennsylvania, draws a corollary between these AI tools and voice recognition or digital PACS in that, in 10 years, we may struggle to think of a time before these tools were commonplace in radiology practices. He contends that while interviewing for residency and fellowship, students ought to inquire about which AI tools — both diagnostic and non-interpretive — departments have adopted. Residents will also increasingly be expected to become facile with these tools in their daily work.

AI in radiology may still be in its infancy and results may be mixed, but research has been accelerating, and a variety of products have already been developed for market. Current and future radiologists have the distinct opportunity to introduce these tools to healthcare settings and use them to improve efficiency of care delivery.

  1. https://www.theverge.com/2022/4/5/23011291/imaging-ai-autonomous-chest-xray-eu-fda.
  2. Richardson M.L., Garwood E.R., Lee Y., Li M.D., Lo H.S., Nagaraju A., Nguyen X.V., Probyn L., Rajiah P., Sin J., Wasnik A.P., Xu K. Noninterpretive Uses of Artificial Intelligence in Radiology. Acad Radiol. 2021 Sep;28(9):1225-1235. doi: 10.1016/j.acra.2020.01.012. Epub 2020 Feb 12. PMID: 32059956.
  3. Jiang D., Dou W., Vosters L., Xu X., Sun Y., Tan T. Denoising of 3D magnetic resonance images with multi-channel residual learning of convolutional neural network. Jpn J Radiol. 2018 Sep;36(9):566-574. doi: 10.1007/s11604-018-0758-8. Epub 2018 Jul 7. PMID: 29982919.
  4. Chen H., Zhang Y., Kalra M.K., Lin F., Chen Y., Liao P., Zhou J., Wang G. Low-Dose CT With a Residual Encoder-Decoder Convolutional Neural Network. IEEE Trans Med Imaging. 2017 Dec;36(12):2524-2535. doi: 10.1109/TMI.2017.2715284. Epub 2017 Jun 13. PMID: 28622671; PMCID: PMC5727581.
  5. Xiang L., Qiao Y., Nie D., An L., Wang Q., Shen D. Deep Auto-context Convolutional Neural Networks for Standard-Dose PET Image Estimation from Low-Dose PET/MRI. Neurocomputing. 2017 Dec 6;267:406-416. doi: 10.1016/j.neucom.2017.06.048. Epub 2017 Jun 29. PMID: 29217875; PMCID: PMC5714510.
  6. Brown A.D., Marotta T.R. A Natural Language Processing-based Model to Automate MRI Brain Protocol Selection and Prioritization. Acad Radiol. 2017 Feb;24(2):160-166. doi: 10.1016/j.acra.2016.09.013. Epub 2016 Nov 23. PMID: 27889399.
  7. Trivedi H., Mesterhazy J., Laguna B., Vu T., Sohn J.H. Automatic Determination of the Need for Intravenous Contrast in Musculoskeletal MRI Examinations Using IBM Watson's Natural Language Processing Algorithm. J Digit Imaging. 2018 Apr;31(2):245-251. doi: 10.1007/s10278-017-0021-3. PMID: 28924815; PMCID: PMC5873465.
  8. Wong T.T., Kazam J.K., Rasiej M.J. Effect of Analytics-Driven Worklists on Musculoskeletal MRI Interpretation Times in an Academic Setting. AJR Am J Roentgenol. 2019 Feb 26:1-5. doi: 10.2214/AJR.18.20434. Epub ahead of print. PMID: 30807228.