Screening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI): a randomised, controlled, parallel-group, non-inferiority, single-blinded, screening accuracy study.

BACKGROUND

Emerging evidence suggests that artificial intelligence (AI) can increase cancer detection in mammography screening while reducing screen-reading workload, but further understanding of the clinical impact is needed.

METHODS

In this randomised, controlled, parallel-group, non-inferiority, single-blinded, screening-accuracy study, done within the Swedish national screening programme, women recruited at four screening sites in southwest Sweden (Malmö, Lund, Landskrona, and Trelleborg) who were eligible for mammography screening were randomly allocated (1:1) to AI-supported screening or standard double reading. The AI system (Transpara version 1.7.0 ScreenPoint Medical, Nijmegen, Netherlands) was used to triage screening examinations to single or double reading and as detection support highlighting suspicious findings. This is a protocol-defined analysis of the secondary outcome measures of recall, cancer detection, false-positive rates, positive predictive value of recall, type and stage of cancer detected, and screen-reading workload. This trial is registered at ClinicalTrials.gov, NCT04838756 and is closed to accrual.

FINDINGS

Between April 12, 2021, and Dec 7, 2022, 105 934 women were randomly assigned to the intervention or control group. 19 women were excluded from the analysis. The median age was 53·7 years (IQR 46·5-63·2). AI-supported screening among 53 043 participants resulted in 338 detected cancers and 1110 recalls. Standard screening among 52 872 participants resulted in 262 detected cancers and 1027 recalls. Cancer-detection rates were 6·4 per 1000 (95% CI 5·7-7·1) screened participants in the intervention group and 5·0 per 1000 (4·4-5·6) in the control group, a ratio of 1·29 (95% CI 1·09-1·51; p=0·0021). AI-supported screening resulted in an increased detection of invasive cancers (270 vs 217, a proportion ratio of 1·24 [95% CI 1·04-1·48]), wich were mainly small lymph-node negative cancers (58 more T1, 46 more lymph-node negative, and 21 more non-luminal A). AI-supported screening also resulted in an increased detection of in situ cancers (68 vs 45, a proportion ratio of 1·51 [1·03-2·19]), with about half of the increased detection being high-grade in situ cancer (12 more nuclear grade III, and no increase in nuclear grade I). The recall and false-positive rate were not significantly higher in the intervention group (a ratio of 1·08 [95% CI 0·99-1·17; p=0·084] and 1·01 [0·91-1·11; p=0·92], respectively). The positive predictive value of recall was significantly higher in the intervention group compared with the control group, with a ratio of 1·19 (95% CI 1·04-1·37; p=0·012). There were 61 248 screen readings in the intervention group and 109 692 in the control group, resulting in a 44·2% reduction in the screen-reading workload.

INTERPRETATION

The findings suggest that AI contributes to the early detection of clinically relevant breast cancer and reduces screen-reading workload without increasing false positives.

FUNDING

Swedish Cancer Society, Confederation of Regional Cancer Centres, and Swedish governmental funding for clinical research.