As radiologists, we strive to deliver high-quality images for interpretation while maintaining patient safety, and to deliver accurate, concise reports that will inform patient care. We have improved image quality with advances in technology and attention to optimizing protocols. We have made a stronger commitment to patient safety, comfort, and satisfaction with research, communication, and education about contrast and radiation issues. But when it comes to radiology reports, little has changed over the past century.
The time for standardization is now
There have been intermittent attempts to structure our reports in recent decades. The reasoning is sound: to decrease confusing variability in language and ensure consistent content. But those attempts failed due to the template’s poor aesthetics and rigid nature.
These past failures, in addition to an inherent desire for independence and control, have made radiologists wary of any new system that will force standardization of their work. But the need to standardize report content and terminology — all radiologists using the same term for the same concept, and hopefully the same term that the ordering physician also uses — is more important than ever.
Report templates are familiar to most radiologists. Both generic and disease-specific templates improve consistency and ensure that required content is included. Despite their value, templates are criticized by those who say they are difficult to read, can hinder radiologists’ thought and expression, and add additional time and effort to the reporting process.
Improvements in technology and the popularity of artificial intelligence (AI) have created a perfect environment for educating the radiology community about the importance of using standard terminology and common data elements (CDEs). While the AI hype has cooled, we recognize potential areas where machine learning algorithms will assist us in detection, diagnosis, and workflow. If these algorithms are ever to be trusted clinically, they must have valid ground truth, established by meaningful, reliable labels. One of the biggest obstacles to creation of these machine learning algorithms for computer vision applications is the availability of the large volume of labeled data needed for training. Our radiology reports contain labels applied by experts, but it is costly and difficult to extract this data from prose or semi-structured reports.
Ideal AI input: structured reports and CDEs
Algorithms that can be safely applied to our work and by extension patient care, require data that is accurate, consistent, and reproducible. The use of structure in reports decreases language variability and ensures that required content is included.
Structure can be applied with report templates, such as those available from RadLex , or specialized systems such as the ACR RADS , which include categorization and management recommendations. CDE macros are modules that can be inserted into prose reports, allowing structure and individualized radiologist description in a hybrid report. Individual CDEs, found at RadElement , can be used in structured or prose reports. All of these structures improve communication and add value to the radiology report. And, all provide an ideal input for machine learning algorithm training.
While report structure is valuable, standardization of language is crucial. As an example, “small vessel ischemic change” seen on an MRI of the brain might be described using that phrase or by a number of synonyms, such as “microvascular disease,” “white matter disease,” “leukoariosis,” or “periventricular ischemic change.” Whether the audience is a referring physician, patient, radiology trainee, or machine-learning algorithm, it is clear that one standard term to describe the finding is easier to understand and learn than five. Adoption of standard language, a constrained vocabulary, is vital to improving reports and creating valid, multipurpose data.
How can we get there from here?
CDE macros can instill necessary structure into our reports. These are a structured set of concepts designed to be inserted into a traditional prose report. The macros contain three to five of the most important pieces of information that should be included in any report addressing the described entity. A CDE is a question with a constrained set of expected responses. The responses may be descriptors, such as “mild, “moderate,” or “severe” in the case of describing spinal canal stenosis; numbers in the case of ASPECTS scoring for stroke; or dimensions in measuring a mass. The key point is that the inclusion of the question ensures report completeness, and the controlled responses ensure consistency in language.
In 2017, ACR, RSNA, and ASNR together created a CDE workgroup to develop CDE macros for the most common neuroradiology examinations. The history and rationale of this project can be found on the ASNR website . Content experts scoured the literature to produce macros based on the most current guidelines, expert consensus, and traditional teachings.
One of the first CDE macros created by the ACR/RSNA/ASNR CDE Workgroup is used in spine trauma . This macro contains individual CDE “questions,” including level of fracture, alignment, and involvement of the posterior elements. The answers are standardized, so all radiologists have the same list to choose from — so there is no variability in the report. The questions and answers in the CDE set are based on the most current spine surgery literature, enabling radiologists to create a report with consistent, standard findings and impressions in the language of the treatment team. This optimizes clinical communication and patient care while highlighting the value of the radiologist to the entire healthcare team. In addition, it provides meaningful, accurate labels that can be easily extracted to train machine-learning algorithms.
The true beauty of this system is that it saves the radiologist time and effort while improving value — an ideal scenario. The macro is inserted into the prose report, and the radiologist answers the multiple-choice questions and selects the appropriate recommendation based on the constrained set of answers. The radiologist does not have to memorize or look up current guidelines; they are provided in the macro. This is also invaluable for trainees, who can use the macros to learn what key elements must be addressed when reporting on a specific clinical question.
Simply put, the use of CDEs and CDE macros decreases inter-reader variability in description and interpretation — and can increase reporting speed, consistency, accuracy, and completeness.
Building the radiology report of the future
Several societies and organizations are now working to build the framework and content needed to standardize our language and reports. The ACR/RSNA/ASNR CDE Workgroup continues to create neuroradiology CDE macros for the most commonly encountered pathologies. Similarly, the ACR RADs library continues to grow, and ACR Assist™ promises to deliver modular structured content that can be incorporated into a conventional narrative report. Reporting systems can use natural language processing to determine when to provide the radiologist with a particular piece of content, such as the ability to select an ACR RADS classification.
A number of societies are also working to establish grading systems and categorization schema. Collaboration will allow us to reach this necessary goal more quickly. As clinical care becomes more data-driven and patient management becomes more algorithmic, so too should radiology reporting — in order to improve patient care with more accurate, actionable, valuable reports.
The key challenge is giving radiologists a strong incentive to change the way they work. It must be clear that this will make our lives easier, allowing us to produce high-value reports with increased speed and efficiency. In addition, improving reports will magnify our role and increase our standing with medical partners and within the healthcare system as a whole. The popularity of AI has prompted many to understand and contribute to the creation of systems that will enable algorithm development and integration into our work. The use of standardized language, CDEs, macros, and templates will facilitate this future.
Wende N. Gibbs, MD | Department of Radiology, Neuroradiology Division | Senior Associate Consultant, Mayo Clinic