fbpx
Call us toll-free: 800-878-7828 — Monday - Friday — 8AM - 5PM EST
Call us toll-free: 800-878-7828 — Monday - Friday — 8AM - 5PM EST

Radiology departments leading the way in artificial intelligence

Radiology departments leading the way in artificial intelligence
February 3, 2022 Rachel Tirabassi

By Sarah Sarvis Milla, M.D., FAAP, and Hansel J. Otero, M.D., FAAP for AAP News

Artificial intelligence (AI) has become a buzzword in many health care environments and even more so in technology-reliant subspecialties such as radiology.

Most radiology departments already deploy complex software to manage increasingly large databases and a completely digital workflow. While most radiologists acknowledge AI’s potential to transform the way medical imaging is practiced, implementation of applications requires radiologists’ expertise. Specifically, advocacy from pediatricians and pediatric radiologists in support of pediatric research is important to assure the safety of AI techniques for use in pediatric patients, as many models are designed using adult data.

Development of AI

AI is computer systems/software that can perform tasks that normally require human intelligence, such as decision-making and visual perception. Approximately 30% of radiologists are using AI to enhance interpretation — most commonly detection of intracranial hemorrhage, pulmonary emboli and mammographic abnormalities, according to the 2020 American College of Radiology Data Science Institute Artificial Intelligence Survey (Allen B, et al. J Am Coll Radiol. 2021;18:1153-1159).

In the near future, AI may be even more useful to radiologists by completing essential tasks such as 3D organ segmentation or tumor measurements, which will improve speed and consistency of the measurements. In pediatrics, AI algorithms have been developed to perform bone measurements in leg length discrepancy studies and bone ages.

AI also can be referred to as machine learning and deep learning, which are the most common methods used to develop AI.

Machine learning describes the traditional process in which data are extracted and labeled/classified to “train” the algorithm. Large datasets are required to achieve accurate output.

Deep learning refers to a multilayered algorithm (“neural networks”) that can extract data automatically and is responsible for most AI applications, including chatbots and facial recognition. Deep learning, however, does not obviate the need for the radiologists. They still are necessary to help navigate uncertainty and weigh in on diagnostic and management possibilities beyond a primary abnormal interpretation.

Benefits of AI

Regardless of the methodology behind them, AI solutions, even if imperfect, offer myriad benefits for clinicians and radiologists. For example, clinicians can use AI to choose the most appropriate imaging test for a patient.

AI can be utilized to optimize radiologists’ workflow by helping them triage and select studies that may require immediate attention such as head CT scans with findings of hemorrhage or stroke identified by a computer algorithm.

AI also can point out areas of potential abnormalities that the radiologist should evaluate carefully, which is particularly important in children who can have normal variants or technique-related findings that can be confused with disease.

Utilizing AI technology can free up radiologists’ time for consultation with clinical teams and/or patients and families.

AI also can help ameliorate human resources shortages in places with limited access to pediatric imaging expertise (see images at the top of the story).

Room for improvement in pediatrics

Despite these developments, AI imaging applications still require deployment under the expertise of radiologists. Moreover, it is uncertain who will be responsible for potential errors if AI were fully automated.

The inappropriate deployment and adoption of models meant for adults represent risks for appropriate and timely care in children. Almost all AI models are trained with adult imaging/data. Therefore, actual performance in children is unknown. Moreover, AI solutions are not required to list the ages of subjects used in testing, training or validation, and AI models approved by the Food and Drug Administration are not required to be labeled for use in adults and/or children.

Pediatricians can expect a rapidly increasing menu of AI solutions in their hospitals, practices and organizations. They should be aware of AI technologies that have not been developed and/or trained and tested in children. In addition, advocacy is needed to develop regulatory standards for appropriate study and use in children.

As early adopters, the pediatric radiology community’s top priorities are AI regulatory clearance and labeling.