Combining ensemble artificial intelligence (AI) models with reads from breast radiologists of mixed experience levels can help health systems consistently diagnose malignant architectural distortion on mammography.
Academic researchers at China’s Guangzhou University, along with industry researchers in the U.S. and China, found as much when they conducted a retrospective review of mammograms acquired and interpreted over a six-year period at the university’s affiliated medical center.
The study is current in Frontiers in Oncology.
In introducing their findings, Daniel Q. Chen, Bo Liu and colleagues name breast architectural distortion as “the most difficult type of tumor to detect and the most commonly missed abnormality due to its inherent subtlety and varying attributes.”
The study’s imaging dataset included 177 malignant and 90 benign architectural distortion findings, all of which were confirmed by pathology results in patients’ electronic medical records.
Comparing the performances of unaided radiologists with AI-aided peers at finding the malignancies, and of AI ensembles alone, the team recorded several telling results.
For example, in one round, combining AI with a junior radiologist brought a specificity of 72.7% and sensitivity of 91.7%. These scores represented significant improvements over AI alone and radiologist alone.
The cumulative results of these and other rounds of comparison “underscore the potential of using deep learning methods to enhance the overall accuracy of pretest mammography for malignant architectural distortion,” the authors comment in their discussion.
They further note the “large gap” in the diagnostic acumen of radiologists across the many regions of China, suggesting the implications of the study extend beyond architectural distortion patients:
Radiologists [in China] are required to report on X-ray, CT and MR examination results, and even imaging technicians are required for this work some of the time. Our results suggest that adding AI to clinical mammography interpretation in settings with junior radiologists could yield significant performance improvements, with the potential to reduce health care system expenditures, address the recurring shortage of experienced radiologists, and reduce missed detection of early breast cancer.”
The study is available in full for free.
More Coverage of Mammogram Architectural Distortion:
1. Yun Wan, Yunfei Tong, Yuanyuan Liu, Yan Huang, Guoyan Yao, Daniel Q. Chen and Bo Liu, “Evaluation of the Combination of Artificial Intelligence and Radiologist Assessments to Interpret Malignant Architectural Distortion on Mammography.” Frontiers in Oncology, April 20, 2022. DOI: https://doi.org/10.3389/fonc.2022.880150.