Radiology practice will be altered by the coming of artificial intelligence, and the process of learning in radiology will be similarly affected. In the short term, radiologists will need to understand the first wave of artificially intelligent tools, how they can help them improve their practice, and be able to effectively supervise their use. Radiology training programs will need to develop curricula to help trainees acquire the knowledge to carry out this new supervisory duty of radiologists. In the longer term, artificially intelligent software assistants could have a transformative effect on the training of residents and fellows, and offer new opportunities to bring learning into the ongoing practice of attending radiologists.
Introduction
The machine learning (ML) technique offers new abilities to create artificially intelligent (AI) software tools capable of autonomously finding patterns in large datasets and underlies many large-scale software products such as Google Translate, Alexa, and Facebook . Researchers have succeeded in applying these techniques to medical applications, with recent successes in detecting diabetic retinopathy , identifying malignant melanomas , and detecting large vessel occlusion in stroke . Although there has been much discussion in the lay press about the role of ML-based AI tools in radiology—including proclamations that we should stop training radiologists —both AI/ML and medical imaging experts predict that these new software tools will be central to radiologists’ practice across the research, clinical, and education domains .
Two large domains of software tools using ML techniques are likely to begin to affect the practice of radiology. The first domain consists of computer vision AI systems, which will likely perform three main tasks within medical imaging: classification, segmentation, and extraction of new biomarkers from raw image data. Radiographic bone age and brain hemorrhage detection networks are examples of classification problems that have been successfully approached with machine learning tools. ML-trained tools for segmentation tasks try to extract a region of interest automatically, such as the left ventricular cavity or fat, muscle, and bone in body composition analysis . AI software also allow us to process images in ways that would be infeasible for humans, such as deriving dual-energy x-ray absorptiometry (DEXA) scores from routine clinical computed tomography (CT) examinations or calculating organ-specific radiation dose estimates . The second domain is natural language processing (NLP)—the ability of software tools to understand human language. The importance of written and spoken language embedded within the practice of radiology suggests another avenue where radiologists’ work will be affected by these new technologies, much as voice recognition (VR) technology has transformed the process of radiology report creation over the past 2 decades.
Our purpose is to suggest paradigms for how radiologists should approach these new tools as they are developed and are deployed across the clinical enterprise, paying special attention to the potential short-term and long-term effects of machine learning on radiology education.
AI Tools as New Pulse Sequences
The successful deployment of ML tools will require integration across image acquisition, archival, and interpretation. Beyond the creation of the algorithms that can perform basic image analysis tasks, these tools must be selectively integrated into the clinical workflow so that radiologists can leverage their power to improve the clinical care of patients. This means that the prospective users—radiologists—will need to be educated on algorithm capabilities and shortcomings. Successful integration will require radiologists to provide input at all levels of the imaging chain and will be similar to the clinical deployment of a new scanner modality. We propose three levels of knowledge that radiologists will need, which are akin to the roles involved in deploying magnetic resonance technology.
The most in-depth knowledge base will be that of the AI tool creators. These physician-scientists will have strong computer science and data science backgrounds to create and to evaluate neural networks and other software tools, and apply those techniques to clinical problems, much as magnetic resonance (MR) researchers design new pulse sequences. The next level of understanding will be the AI tool deployers. These are the radiologists who may not have the technical capability to create new tools, but understand the core concepts well enough to know which systems will interact well with their scanner fleet and patient population, allowing them to create appropriate “protocols” for assisted or autonomous image interpretation. They will require a new skill set for the evaluation of AI tools, both when considering new tools and when evaluating the ongoing performance of deployed tools.
Get Radiology Tree app to read full this article<
AI in Academic Departments in the Short Term
Get Radiology Tree app to read full this article<
Get Radiology Tree app to read full this article<
Get Radiology Tree app to read full this article<
Get Radiology Tree app to read full this article<
Get Radiology Tree app to read full this article<
Get Radiology Tree app to read full this article<
Longer Term Effects of AI on Radiology Education
Get Radiology Tree app to read full this article<
Get Radiology Tree app to read full this article<
Get Radiology Tree app to read full this article<
Get Radiology Tree app to read full this article<
Get Radiology Tree app to read full this article<
Get Radiology Tree app to read full this article<
AI-enhanced Lifelong Learning
Get Radiology Tree app to read full this article<
Get Radiology Tree app to read full this article<
Conclusions
Get Radiology Tree app to read full this article<
References
1. Thrall J.H., Li X., Li Q., et. al.: Artificial intelligence and machine learning in radiology: opportunities, challenges, pitfalls, and criteria for success. J Am Coll Radiol 2018;
2. Lakhani P., Prater A.B., Hutson R.K., et. al.: Machine learning in radiology: applications beyond image interpretation. J Am Coll Radiol 2018; 15: pp. 350-359.
3. Erickson B.J., Korfiatis P., Akkus Z., et. al.: Machine learning for medical imaging. Radiographics 2017; 37: pp. 505-515.
4. Gulshan V., Peng L., Coram M., et. al.: Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA 2016; 316: pp. 2402-2410.
5. Esteva A., Kuprel B., Novoa R.A., et. al.: Dermatologist-level classification of skin cancer with deep neural networks. Nature 2017; 542: pp. 115-118.
6. Viz.ai. Viz.ai. Available at https://www.viz.ai/
7. Press Announcements—FDA permits marketing of clinical decision support software for alerting providers of a potential stroke in patients. Available at https://www.fda.gov/NewsEvents/Newsroom/PressAnnouncements/ucm596575.htm
8. Mukherjee S.: A.I. versus M.D. New Yorker 2017; Available at https://www.newyorker.com/magazine/2017/04/03/ai-versus-md
9. Lee H., Tajmir S., Lee J., et. al.: Fully automated deep learning system for bone age assessment. J Digit Imaging 2017; 30: pp. 427-441.
10. Larson D.B., Chen M.C., Lungren M.P., et. al.: Performance of a deep-learning neural network model in assessing skeletal maturity on pediatric hand radiographs. Radiology 2017; 170236:
11. Grewal M., Srivastava M.M., Kumar P., et. al.: RADNET: radiologist level accuracy using deep learning for hemorrhage detection in CT scans. arXiv [csCV]; Available at http://arxiv.org/abs/1710.04934
12. Desai V., Flanders A.E., Lakhani P.: Application of Deep Learning in Neuroradiology: Automated Detection of Basal Ganglia Hemorrhage using 2D-Convolutional Neural Networks. arXiv [csCV]; Available at http://arxiv.org/abs/1710.03823
13. Prevedello L.M., Erdal B.S., Ryu J.L., et. al.: Automated critical test findings identification and online notification system using artificial intelligence in imaging. Radiology 2017; 285: pp. 923-931.
14. Cardio—Arterys. Available at https://arterys.com/cardio
15. Arterys Cardio 510(k) Premarket Notification. Available at https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K163253
16. Lee H., Troschel F.M., Tajmir S., et. al.: Pixel-level deep segmentation: artificial intelligence quantifies muscle on computed tomography for body morphometric analysis. J Digit Imaging 2017; 30: pp. 487-498.
17. Bregman-Amitai O., Elnekave E.: Systems and methods for emulating DEXA scores based on ct images. US Pat 2016; Available at https://patentimages.storage.googleapis.com/b7/fa/53/724bd0e74dbeac/US20160015347A1.pdf
18. Cho J., Lee E., Lee H., et. al.: Machine Learning Powered Automatic Organ Classification for Patient Specific Organ Dose Estimation. In: Proceedings of the Society for Imaging Informatics in Medicine Annual Meeting; Available at https://c.ymcdn.com/sites/siim.org/resource/resmgr/siim2017/abstracts/analytics1-Do.pdf
19. RADLogics : http://radlogics.com/
20. ACR Data Science Institute : http://www.acrdsi.org/