MASAI Trial: AI-assisted Mammography Screening — Clinical Safety & Workload (Bullet Notes)

MASAI trial evaluates AI-supported screen-reading versus standard double reading in population-based mammography screening.

Type: randomised, non-inferiority, single-blinded, parallel design.
Population: women aged 40–80 at four Swedish screening sites; includes general screening intervals of 1.5–2 years and annual screening for moderate hereditary risk or history of breast cancer.
Randomisation: 1:1 allocation via PACS; automated pseudo-randomization; participants and radiographers masked to allocation; radiologists reading not masked.
Interventions: AI triage with Transpara v1.7.0: exams with risk scores 1–9 → single reading; score 10 → double reading; CAD marks for scores 8–10; radiologists have AI risk scores and CAD marks at reading time.
Control: standard unassisted double reading.
Outcomes: prespecified safety analysis focusing on early screening performance (cancer detection rate, recall rate, false positive rate, PPV of recall, cancer type) and screen-reading workload; primary endpoint is interval cancer rate to be assessed after 2 years in a full 100,000-participant cohort.
Analysis populations: modified intention-to-treat (mITT); exclusions include technical recalls and non-suspicious lymphoma cases.

Enrollment and analysis: 80,033 randomized (AI: 40,003; control: 40,030); 39,996 AI and 40,024 control included in the clinical safety analysis; 13 excluded.
Cancer detection rate: AI $6.1\times10^{-3}$ per participant; control $5.1\times10^{-3}$ per participant; ratio $RR=1.2$ ; $p=0.052$ .
Recall rate: AI $0.022$ ; control $0.020$ (i.e., 2.2% vs 2.0%).
False positive rate: AI $0.015$ ; control $0.015$ (1.5% in both).
Positive predictive value (PPV) of recall: AI $0.283$ ; control $0.248$ .
Cancer type among detections: AI invasive $0.75$ of cancers; in situ $0.25$ ; control invasive $0.81$ ; in situ $0.19$ .
Screen-reading workload: AI readings $46{,}345$ vs control $83{,}231$ ; reduction $44.3\%$ ; total of $36{,}886$ fewer readings.
Interpretation: AI-supported screening detected more cancers with similar false-positive burden and substantially reduced workload; primary endpoint interval cancer rate to be evaluated in full cohort.

High-risk group (risk score 10): cancer detection rate $72.3\text{ per }1000$ ; 208 of 2875 participants with cancers in this subgroup.
Extra-high-risk group (top 1%, marked as 10H): recalls $189$ ; cancers detected $136$ ; PPV of recall $72.0\%$ ; detection rate $277.6\text{ per }1000$ in this category; this group contained $55.7\%$ of all screen-detected cancers.
Single-reading with AI (risk scores 1–9): $92.0\%$ of exams; cancers detected by single-reading with AI: $36$ (14.8\% of all cancers detected).

Strengths: real-world screening setting; minimal exclusions; large sample; integration with national programme.
Limitations: single-centre study; one mammography device and one AI system; radiologist experience may limit generalisability; potential automation bias considerations; true false-positive rate will depend on follow-up for interval cancers.

AI-assisted triage plus detection can be considered safe and substantially reduces screen-reading workload without increasing recalls or false positives.
Primary efficacy endpoint (interval cancer rate) pending in full 100,000-participant cohort with 2-year follow-up.
Future implementation requires careful strategy (which readings use AI data, thresholds, and monitoring of algorithm performance).

AI-supported mammography screening demonstrated a similar cancer detection rate to standard double reading with a large reduction in screen-reading workload, supporting potential clinical implementation while awaiting full interval-cancer-rate data.