Using Artificial Intelligence to Improve Radiology Workflow

Julie Song, MPH, CPHRM, Patient Safety Risk Manager II, The Doctors Company

The success of artificial intelligence (AI) in research settings has been difficult to replicate in routine clinical practice. The initial concern that AI might replace radiologists has been tempered by the logistical, ethical, and legal hurdles encountered with AI implementation. A 2020 survey by the American College of Radiology revealed that approximately 30 percent of radiologists were using AI in their practices. While this statistic shows a modest penetration into the field, it is far from the “rise of the machines” takeover once feared.

While the ethical and legal challenges around AI are nuanced and will take time to resolve, AI may achieve a faster adoption rate by enhancing the efficiency of radiology workflow in areas such as ordering studies, determining scan protocols, acquiring images, interpreting images, and generating reports. It is, however, also important to recognize that AI applications in these areas, while having an impact on improving patient safety, may also have downstream consequences that are not yet fully realized.

Consider the following potential AI workflow improvements.

Examination Planning and Ordering Studies

With the ever-growing amount of patient information in the EHR, it may be difficult for a radiologist to synthesize all of the data into a customized imaging plan. This is where AI’s ability to quickly scan and collate information into a meaningful plan based on the patient’s clinical indication could be leveraged.1 By optimizing the process of ordering studies, AI can reduce the number of unnecessary scans.

Time saved by the radiologist could be better invested in patient-facing areas, such as educating the patient about the procedure or exam, conducting informed consent, and answering questions and concerns that could result in improved patient satisfaction.

“Machine learning,” a subset of AI, is based on the idea that the algorithm becomes more accurate in problem solving through experience and data. Although the algorithm is not explicitly programmed, its exposure to types of data and experience can affect how it solves problems. Through machine learning, AI can be refined to the point that it could decrease the number of imaging studies by accurately predicting the diagnosis from the clinical indications and information gathered in the EHR alone.2

Scan Protocoling

Scan protocoling is an integral and, often, time-consuming part of a radiologist’s workflow. The scan protocol process correlates imaging modality and contrast administration to the clinical indication. The information necessary to arrive at the correct protocol, however, involves reviewing patient charts, relevant labs, and previous studies to identify contraindications when scheduling contrast-enhanced studies. AI can streamline the process by quickly synthesizing the information and recommending the correct imaging protocol.

AI could also evolve into identifying protocols for the most commonly occurring clinical indications with minimal protocol variability so that the radiologist can focus on more complex cases.2 This, combined with a radiology worklist that prioritizes protocols based on the likelihood of a critical finding, could decrease turnaround time. In creating AI-assisted protocols, however, it is important to have a diverse data source to prevent reinforcing any health disparities.

Image Acquisition

AI also has great potential for improving the radiology image acquisition process. To ensure that suspicious findings are diagnosed early, image quality is imperative. Poor quality images require follow-up studies that often result in additional delays in rescheduling the exam and potentially exposing the patient to unnecessary radiation. AI-supported guidance on patient positioning, contrast dosing, or image sequencing not only improves patient safety, but it could also result in cost savings over time.

Image Interpretation

Image interpretation has been the most discussed AI-enhanced portion of the radiology workflow because of its impact on the clinical contribution of the radiologist. Technology advances faster than our legal system, so it remains uncertain how the use of AI will play out in medical-legal cases. Will the attending radiologist be held liable, or will the AI algorithm that the radiologist relied on be liable? What accountability will the company that developed the AI algorithm have—and will that company still be liable when the AI algorithm evolves through machine learning?

Defense attorney Terrence Schafer of Doyle Schafer McMahon LLP suggests the following documentation strategies: “If the radiologist chooses to reject an AI algorithm finding, it is important to document the rationale of the decision to prevent an allegation of disregarding a safeguard that was available to the clinician. Confirm that the radiologist reviewed the area of concern brought to light by the AI algorithm, but document that the clinician in his or her professional judgment disagrees with the AI analysis due to X, Y, and Z reasons. These actions establish the clinician as rendering a carefully considered professional opinion with the advantages of using AI without relying on it exclusively.”

An acceptable method of incorporating AI during image interpretation is having the radiologist initially conduct the interpretation with AI running in the background. Using AI as a backup minimizes any biases that might arise from the AI algorithm while suggesting findings that the radiologist might not have initially identified. Used in this way, AI can be a powerful tool to identify areas that need closer inspection, and it can help capture secondary findings that the radiologist might have missed.

Report Generation

Because radiology reports have potential legal repercussions, AI-assisted reports must be generated with criteria other than simply efficiency. A basic use of an AI-assisted report is integrating relevant information from PACS (picture archiving and communication systems) or the EHR into the designated findings. One example is incorporating the measurements of the patient’s anatomical structure or relevant clinical history for analysis by the radiologist. While information directly from PACS or an EHR is objective, it is important to delineate which portions of the final analysis are independent clinical conclusions and which are the AI-suggested recommendations.


The degree of an organization’s interoperability and standardization will determine the success of its transition to an AI-enabled workflow. Currently, no single standard for AI interoperability exists. On March 3, 2021, the Integrating the Healthcare Enterprise Radiology Technical Committee, in a joint effort between medical imaging societies, industry, and radiology communities, produced a white paper titled AI Interoperability in Imaging. Once finalized, the report will provide a roadmap of the AI interoperability needs and proactively address challenges to adoption of AI in the imaging ecosystem. It is hoped that radiologists will view AI as an asset to their work instead of a threat.

For additional assistance, contact the Department of Patient Safety and Risk Management at (800) 421-2368 or by email.


  1. Hardy M, Harvey H. Artificial intelligence in diagnostic imaging: impact on the radiology profession. 2020;93(1108).
  1. Kapoor N, Lacson R, Khorasani R. Workflow applications of artificial intelligence in radiology and an overview of available tools. J Am Coll Rad. 2020;17(11):1363-1370.

The guidelines suggested here are not rules, do not constitute legal advice, and do not ensure a successful outcome. The ultimate decision regarding the appropriateness of any treatment must be made by each healthcare provider considering the circumstances of the individual situation and in accordance with the laws of the jurisdiction in which the care is rendered.

J13102 10/21