Noninferiority Randomized Trial of AI-Augmented Mammography Reading in Breast Cancer Screening

A randomized, controlled clinical trial for AI-supported mammography readings demonstrated that AI reads of mammogram scans led to fewer interval breast cancer diagnoses than with standard double reads by radiologists. Findings from the MASAI trial were published in The Lancet.

“Our study is the first randomized controlled trial investigating the use of AI in breast cancer screening and the largest to date looking at AI use in cancer screening in general,” said corresponding study author Kristina Lång, PhD, Division of Diagnostic Radiology, Department of Translational Medicine, Lund University, Sweden. “It finds that AI-supported screening improves the early detection of clinically relevant breast cancers, which led to fewer aggressive or advanced cancers diagnosed in between screenings.”

Researchers in Sweden conducted the randomized, controlled, non-inferiority, single-blinded, population-based screening accuracy MASAI clinical trial to look at the impact of AI-supported readings of mammography screenings on the rate of interval cancers, or primary breast cancers diagnosed between rounds of screenings. Patients were equally randomly assigned into an interventional group of patients receiving AI-supported mammography screening or a control group of patients receiving standard double readings without AI support.

The primary outcome measure was the interval cancer rate, and secondary endpoints included interval cancer characteristics, sensitivity, specificity, and sensitivity according to the patient's age, breast density, and cancer type.

Interim safety results of the study were previously published in The Lancet Oncology and showed that the use of AI support reduced radiologists' workload by 44%. Another analysis of the trial showed that 29% more cancers were detected with AI-supported readings without increasing the rate of false positives.

Results

A total of 105,934 women were randomly assigned to one of the two arms, but 19 patients were excluded from the final analysis.

The interval cancer rate in the intervention group was 1.55 (95% confidence interval [CI] = 1.23–1.92) per 1,000 participants and 1.76 (95% CI = 1.42–2.15) per 1,000 participants in the control group. The proportion ratio for noninferiority was 0.88 (95% CI = 0.65–1.18; P = .41). In the interventional group, there were fewer interval cancers diagnosed with unfavorable characteristics, including invasive disease (75 vs 89 in the control group), T2 or higher stage (38 vs 48), and non-luminal A disease (43 vs 59).

Sensitivity was higher in the interventional group than the control group (80.5% vs 73.8%; P = .031), but specificity was the same for both groups (98.5; 95% CI = 98.4%–98.6%; P = .88). The sensitivity was consistently higher in the interventional group than the control group by factors such as age, breast density, and invasive cancer, but not for in-situ cancer.

“Our study does not support replacing health-care professionals with AI as the AI-supported mammography screening still requires at least one human radiologist to perform the screen reading, but with support from AI,” said first study Jessie Gommers, MSc, a PhD student at Radboud University Medical Centre in the Netherlands. “However, our results potentially justify using AI to ease the substantial pressure on radiologists’ workloads, enabling these experts to focus on other clinical tasks, which might shorten the waiting times for patients.”

Model Methods

Artificial intelligence was used to triage if a scan required a single or double reading by radiologists. The AI system (Transpara, version 1.7.0) also functioned as a detection-support tool by highlighting suspicious findings on mammogram scans. Transpara was trained, validated, and tested on over 200,000 mammography scans from more than 10 countries to account for diverse populations and image types.

In the study, the AI system analyzed mammograms and assigned an overall risk score ranging from one to 10 for suspicious findings. Any mammogram assigned a score of 10 necessitated a double reading, whereas all other scores were assigned a single reading.

As detection support, the system marked suspicious calcifications and soft-tissue lesions by assigning region-specific scores ranging from one to 98. Participating radiologists were instructed to review these marks only after completing their independent reading.

DISCLOSURE: The study was funded by the Swedish Cancer Society, Confederation of Regional Cancer Centres, and Swedish governmental funding for clinical research. For full disclosures of the study authors, visit thelancet.com.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.