4 areas where AI is having its biggest impact on breast imaging
Artificial intelligence (AI) technologies are advancing at a rapid rate and starting to make a direct impact on breast imaging. There is still a lot of work to be done, however, before AI can truly be trusted with making decisions that may impact a patient’s survival, according to a new commentary published in the American Journal of Roentgenology.
Author Ellen B. Mendelson, MD, Feinberg School of Medicine at Northwestern University in Chicago, studied the National Library of Medicine and PubMed for original research on AI from 1993 to 2018, with an emphasis on the last 10 years. She also reviewed recent articles from the New York Times and the New Yorker for additional insights.
During her research, Mendelson explored how AI is currently impacting breast imaging and noted both key advances and key limitations . These are four specific areas she focused on in her commentary:
1. Diagnosis
If specialists are expecting AI to provide a definitive diagnosis without a human physician overseeing the results, Mendelson warned, they are in for a rude awakening. Radiologists and other physicians are still needed to confirm the diagnosis due to the possibility of an error.
“If AI incorrectly analyzes an examination, resulting in an improper management recommendation, incorrect report, and miscommunication to the patient, the error first must be recognized by the interpreting radiologist, it is hoped, or by the referring physician, the technologist, the patient, or someone else,” she wrote. “These are important human interactions with computer behavior.”
This does not mean AI should be ignored altogether or tossed to the side, Mendelson added, but it is something radiologists must keep in mind.
She also noted that AI’s ability to analyze images for various features is “an exciting prospect for research in diagnosis,” noting that they can help detect hard-to-see lesions and decrease false-negative interpretations.
2. Clinical Decision Making
AI can also help specialists with clinical decision making, but like diagnosis, the technology is still far from perfect and requires physician oversight.
“Feature analysis and pattern recognition for breast imaging were codified in the ACR's BI-RADS lexicons, and BI-RADS assessments have been incorporated in AI data as measures of the likelihood of malignancy internationally in breast imaging studies,” Mendelson wrote. “Nevertheless, uncertainty in diagnosis remains, with false-positives the result of a lack of interpretive confidence, poor images to interpret, overlap in diagnostic criteria, or even liability fears”
She also explored work in machine learning (ML) that has found artificial neural networks (ANNs) can help standardize mammographic interpretation.
“Without established methods to detect output errors of ML algorithms, it is unlikely that breast imagers will trust AI algorithms in areas where misdiagnosis can be catastrophic, but it is even more important that they work together with the computer scientists, statisticians, and information technologists in adapting AI to the needs of breast imaging,” Mendelson added.
3. Patient Outcomes
While AI is now able to help generate outcome analyses, ANNs and deep learning algorithms “have the potential to improve screening outcomes by increasing predictability, not only of findings representing cancer, but with radiomic yoking radiology and pathology images and additional data, additional features and their previously unrecognized pathologic correlates may also sharpen the decision between recall and routine management.”
AI also offers radiologists a valuable tool for improving patient management, Mendelson wrote, and algorithms could be used to “settle a management debate,” although they “lack the interchange of ideas and experience” of an actual, experienced breast imaging specialist.
4. Workflow
One of the greatest ways AI can provide value to today’s radiologists is through improving day-to-day workflow and allowing them to be more efficient. Less time spent on more mundane tasks equals more time spent reading studies and helping patients.
Natural language processing (NLP), for instance, can help breast imagers with such tasks as automated image segmentation with labeling, measurement and comparing newer studies with prior studies, Mendelson explained.
Overall, Mendelson sees “great potential for AI” to help breast imagers in all of these areas. The technologies are still evolving, however, and as long as “algorithmic error” is still a possibility, AI “cannot and should not be relied on” for making decisions that may affect a patient’s survival.
Michael has more than 18 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.