Screening Mammograms: Is AI Better Than a Radiologist?

Andrew M. Kaunitz, MD


January 31, 2020

This transcript has been edited for clarity.

The accuracy of screening mammography varies substantially, even among expert radiologists.

Artificial intelligence (AI) has demonstrated promising results when used in a variety of medical imaging applications. To assess AI's utility in interpreting screening mammograms, investigators supported by Google used British and US clinical datasets to create a so-called "deep learning model." The goal was to identify biopsy-confirmed cases of breast cancer diagnosed subsequent to screening mammograms performed in British and US women.

Then they compared the performance of their AI-based system with that of clinical radiologists. In Britain, two radiologists routinely interpret each mammogram. In contrast, in the United States, one radiologist reads each mammogram in most cases.

To assess whether AI's usefulness can be generalized, the investigators tested their UK-based AI model using US data.

Finally, the investigators compared the overall performance of their AI system with that of six US radiologists who were not otherwise involved in the study.

Compared with UK and US study radiologists, the AI-based system significantly reduced false positives, particularly among the US radiologists. The sensitivity and specificity of the AI system were both superior to those of the six US readers.

It is important to point out that neither AI nor human radiologist interpretation of mammograms is foolproof. AI identified one cancer case that was missed by all six independent radiologists. However, a second case identified by all of the radiologists was missed by AI.

The findings of this study are encouraging and suggest that AI might help identify screening mammograms requiring extra attention by radiologists.

However, we should remember that while much enthusiasm surrounded computer-aided software when it became available in the 1990s, that software ultimately failed to improve mammogram performance.

Before implementing AI clinically, its usefulness for interpreting screening mammograms should be evaluated in broader populations of patients and radiologists.

Thank you for the honor of your time. I am Andrew Kaunitz.

Dr Andrew Kaunitz is a tenured University of Florida term professor and associate chair of the Department of Obstetrics and Gynecology at the University of Florida College of Medicine-Jacksonville. Dr Kaunitz has published more than 240 articles in peer-reviewed journals, including the New England Journal of Medicine, JAMA, and Obstetrics and Gynecology.

Follow Medscape on Facebook, Twitter, Instagram, and YouTube


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.
Post as: