Improving Decision-Making Time and Accuracy in Radiology Using AI

Artificial intelligence (AI) is frequently alleged to support diagnostic decisions made by radiologists. For example, it appears to support multiple radiological aspects such as augmenting patient outcomes, offering versatile workflows, reducing radiation doses and the need for contrast agents (CAs), and delivering tailored diagnostics (1). However, the question here is whether (or how) AI can improve decision-making time and promote accuracy in radiology.

Overall, computer-aided diagnosis (CAD) can shorten reading time and enhance outcomes for many diseases. In recent research, for example, AI was found to support radiologists in screening mammographic examinations of 240 women for potential breast cancer, with sensitivity (86% vs 83%) and specificity (79% vs 77%) in reading images that were improved with AI support, compared to unaided reading (2). Elsewhere in Nepal, deep learning (DL) systems (e.g., qXR, INSIGHT, CAD4TB, etc.) were employed to interpret chest radiographs (CXRs) for potential abnormalities caused by pulmonary tuberculosis (TB) in 1196 individuals. In this research, the specificity (≥80%) and sensitivity (≥95%) of DL systems (qXR, CAD4TB, and Lunit) was higher than that of Nepali radiologists in interpreting CXRs. Accordingly, these DL systems outperformed radiologists in labeling patients with bacteriologically approved TB compared to healthy individuals (3). These and other findings suggest that AI-based interventions can markedly reduce the need for follow-up tests and augment the capacity of radiologists to diagnose diseases.

Today, AI algorithms (particularly DL) are extensively used to analyze medical images. Examples of AI applications in radiology can be found in oncology (e.g., thoracic imaging, abdominal and pelvic imaging, colonoscopy, mammography, radiation oncology, and brain imaging). Similarly, machine learning models are today trained to efficiently classify patients, thereby supporting clinical decision-making by promoting the detection of abnormalities and then monitoring and characterizing potential changes in these conditions (4). AI-powered programs are automated to best extract radiomic data from radiological images that are not readily detectable by visual inspection. This substantially improves prognosis and diagnostic accuracy (5) and provides verified and trusted platforms in radiology practices (i.e., by enhancing interpretability) (6).

Currently, AI is believed to both alter the interpretation of images (e.g., with higher sensitivity and specificity) and affect all radiology’s clinical practices (e.g., after the advent of DL, supervised learning, machine learning, etc.). Overall, AI can affect radiology by exerting the following effects (7):

• AI serves as a tool to optimize radiology workflow; e.g., by using versatile classifiers to identify abnormalities, and detect stroke and intracranial hemorrhage (on non-contrast brain CT) and acute stroke (on diffusion-weighted MRI).
• AI can remarkably shorten a typical radiology course (i.e., from examinations to interpretations); this is achieved by identifying targets (e.g., abnormalities) from radiography images and then merging them with data available on image metadata to generate interpretable output images.
• AI provides intelligent interpretation (i.e., by alerting the radiologist about potential findings) and reporting (i.e., auto-populated reporting by, for example, NLP (natural language processing) that minimize the time typically spent by a radiologist to report and/or NLP’s ability to extract data from EMRs (electronic medical reports)).

Accordingly, the unique role of AI in promoting radiology outcomes is indisputable. For example, research shows that an AI system (merging DL and Bayesian networks) exhibits an accuracy of 90% in generating differential diagnoses at brain MRI, compared to accuracies of 86% (by academic neuroradiologists), 77% (by neuroradiology fellows), 57% (by general radiologist), and 56% (by radiology residents). Indeed, the Bayesian network (merging clinical information with imaging features) is 85% and 64% accurate in terms of T3DDx (top three differential diagnoses) and TDx (correct top diagnosis) indices, compared to corresponding accuracies of 56% and 36% (for radiology residents) and 53% and 31% (for general radiologists).

However, despite the multitude of reports on the capacity of AI in promoting radiology, the efficacy of AI commercially available products needs to be strictly reflected. For example, research shows that of all the 100 AI products in the current market, only 18 possess (potential) clinical impacts (9), implying that AI-based radiology is yet on its initial journey. There are further problems with deployment approaches and financial issues. When addressing issues with interpretability and performance, AI can assist radiologists in minimizing decision-making time and improving the accuracy of diagnosis (10). Nonetheless, radiologists play a pivotal role in any radiology practice, even with recent breakthroughs in AI, and a successful radiography plan entails radiologists acquiring AI principles and technologies in the future.

References

1. Vn Leeuwen, Kicky G., Maarten de Rooij, Steven Schalekamp, Bram van Ginneken, and Matthieu JCM Rutten. “How does artificial intelligence in radiology improve efficiency and health outcomes?” Pediatric Radiology (2021): 1-7.
2. Rodríguez-Ruiz, Alejandro, Elizabeth Krupinski, Jan-Jurre Mordang, Kathy Schilling, Sylvia H. Heywang-Köbrunner, Ioannis Sechopoulos, and Ritse M. Mann. “Detection of breast cancer with mammography: effect of an artificial intelligence support system.” Radiology 290, no. 2 (2019): 305-314.
3. Qin, Zhi Zhen, Melissa S. Sander, Bishwa Rai, Collins N. Titahong, Santat Sudrungrot, Sylvain N. Laah, Lal Mani Adhikari et al. “Using artificial intelligence to read chest radiographs for tuberculosis detection: A multi-site evaluation of the diagnostic accuracy of three deep learning systems.” Scientific reports 9, no. 1 (2019): 15000.
4. Hosny, Ahmed, Chintan Parmar, John Quackenbush, Lawrence H. Schwartz, and Hugo JWL Aerts. “Artificial intelligence in radiology.” Nature Reviews Cancer 18, no. 8 (2018): 500-510.
5. Thrall, James H., Xiang Li, Quanzheng Li, Cinthia Cruz, Synho Do, Keith Dreyer, and James Brink. “Artificial intelligence and machine learning in radiology: opportunities, challenges, pitfalls, and criteria for success.” Journal of the American College of Radiology 15, no. 3 (2018): 504-508.
6. Reyes, Mauricio, Raphael Meier, Sérgio Pereira, Carlos A. Silva, Fried-Michael Dahlweid, Hendrik von Tengg-Kobligk, Ronald M. Summers, and Roland Wiest. “On the interpretability of artificial intelligence in radiology: challenges and opportunities.” Radiology: artificial intelligence 2, no. 3 (2020): e190043.
7. Syed, A. B., & Zoga, A. C. (2018, November). Artificial intelligence in radiology: current technology and future directions. In Seminars in musculoskeletal radiology (Vol. 22, No. 05, pp. 540-545). Thieme medical publishers.
8. Rauschecker, Andreas M., Jeffrey D. Rudie, Long Xie, Jiancong Wang, Michael Tran Duong, Emmanuel J. Botzolakis, Asha M. Kovalovich et al. “Artificial intelligence system approaching neuroradiologist-level differential diagnosis accuracy at brain MRI.” Radiology 295, no. 3 (2020): 626-637.
9. Van Leeuwen, Kicky G., Steven Schalekamp, Matthieu JCM Rutten, Bram van Ginneken, and Maarten de Rooij. “Artificial intelligence in radiology: 100 commercially available products and their scientific evidence.” European radiology 31 (2021): 3797-3804.
10. Yasaka, Koichiro, and Osamu Abe. “Deep learning and artificial intelligence in radiology: Current applications and future directions.” PLoS medicine 15, no. 11 (2018): e1002707.