Artificial Intelligence (AI), Breast Cancer Screening
Conditions
Keywords
Breast cancer screening, Artificial intelligence (AI)
Brief summary
The goal of this clinical trial is to compare patient-centered outcomes when screening digital breast tomosynthesis (DBT) exams are interpreted with versus without a leading FDA-cleared artificial intelligence (AI) decision-support tool in real-world U.S. settings and to assess patients' and radiologists' perspectives on AI in medicine. The main question it aims to answer is: Does an FDA-cleared AI decision-support tool for digital tomosynthesis (DBT) improve screening outcomes in real world US clinical settings? This trial will include all interpreting radiologists and all adult patients undergoing screening mammography at any of the participating breast imaging facilities across 6 regional health systems (University of California, Los Angeles (UCLA), University of California, San Diego (UCSD), University of Washington-Seattle, University of Wisconsin-Madison, Boston Medical Center, and University of Miami) during the trial period. All screening mammograms at these facilities will be randomized to either intervention (radiologist assisted by an AI decision support tool) versus usual care (radiologist alone) to see if interpreting these mammograms with the AI tool's assistance improves patient screening outcomes. We are targeting 400,000 screening exams across the participating health systems in this trial.
Detailed description
During the RCT the AI support tool will be randomized to be turned on or off (1:1) at the mammography exam level. Patients who return for screening exams in year 2 of recruitment will be randomized again (e.g., they will not retain their prior randomization). Radiologists will not be able to sort exams based on AI availability or AI scores. Randomizing by exam level will ensure that we capture a substantial number of interpretations with vs. without AI for each radiologist, allowing for quantification of the radiologist-level AI learning curve. We are not randomizing at the facility level as some radiologists interpret exams acquired at different facilities on the same day. By randomizing AI at the exam level, we will have the best ability to estimate and adjust for temporal trends in screening outcomes across individual radiologists. Randomization across large regional health systems will be managed independently at each participating site. Our RCT randomizes screening mammography exams to be interpreted either with or without an AI decision-support tool. As a result, radiologists cannot be blinded to study arm during screening mammography interpretation. However, interpreting radiologists and facility staff (e.g., those scheduling the exams) will not know in advance which patients will be randomized to the AI tool. Randomization occurs within minutes after the breast imaging acquisition (i.e., when the mammography technologist captures the images) by an automated system that was developed by a third-party AI platform and successfully piloted at UCLA. Thus, the AI data (or lack thereof) is embedded within the mammogram before the radiologist opens the exam, preventing any option to add AI to an exam randomized to be interpreted without AI. Radiologists will be aware of AI availability only at the time of interpretation, as AI information will appear upon opening the exam (e.g., the AI information pops up with the exam images).
Interventions
DEVICEArtificial intelligence (AI) decision-support tool
The intervention is an AI decision-support tool to help radiologists interpret 3D screening mammograms. For exams randomized to this intervention arm, the first image displayed to the radiologist upon opening an exam on the viewing station will be a one-page, standardized AI report showing the overall exam risk (elevated, intermediate, or low), image region markings, lesion scores from 1-100 (100 being the highest suspicion), bounding boxes, and relevant slice locations for 3D exams. Radiologists can toggle markings on/off and retain full control over the final interpretation of the exam as positive or negative (i.e., they can choose to ignore the AI information). Randomization occurs 1:1 at the exam level via automated code at image acquisition. Returning patients in year two will be re-randomized. Radiologists cannot filter their exam lists by AI availability or risk, and randomization will be independently managed at each participating health system.
Sponsors
University of California, Los Angeles
University of California, San Diego
University of Wisconsin, Madison
Boston Medical Center
Patient-Centered Outcomes Research Institute
University of Washington
California Breast Cancer Research Program
University of Miami
University of California, Davis
Jonsson Comprehensive Cancer Center
Study design
Allocation
RANDOMIZED
Intervention model
PARALLEL
Primary purpose
SCREENING
Masking
SINGLE (Subject)
Intervention model description
This is a study of an FDA-cleared artificial intelligence (AI) decision-support tool.
Eligibility
Sex/Gender
ALL
Age
18 Years to No maximum
Healthy volunteers
Yes
Inclusion criteria
This trial will include all radiologists interpreting screening mammography and all adult patients undergoing screening mammography at any of the participating breast imaging facilities across 6 regional health systems (UCLA, UC San Diego, University of Washington-Seattle, University of Wisconsin-Madison, Boston Medical Center, and University of Miami) during the trial period. Individuals must meet the following eligiblity criteria. Inclusion Criteria: 1. Be at least 18 years of age or older 2. Receive a screening mammogram at one of the participating breast imaging facilities OR be a radiologist who interprets screening mammograms at one of the participating breast imaging facilities.
Exclusion criteria
1\. Patients who have opted out of all research at the health system
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Cancer detection rate | Cancer diagnosed within 90 days of positive study entry screening mammogram | Number of screening exams recommended for breast biopsy (final Breast Imaging- Reporting and Data System \[BI-RADS\] assessment of 4 or 5) resulting in detected cancer, per 1,000 screening exams |
| Recall rate | Through study completion, an average of 1 year | Number of screening exams recalled for diagnostic work-up (initial BI-RADS assessment of 0, 3, 4, or 5), per 1,000 screening exams |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| False positive short-interval follow-up recommendation rate | No cancer diagnosed within 365 days of a positive study entry screening mammogram | Proportion of screening exams recalled for short-interval follow-up (final BI-RADS assessment of 3) with no breast cancer diagnosed within 1 year |
| False positive biopsy recommendation rate | No cancer diagnosed within 365 days of a positive study entry screening mammogram | Proportion of screening exams recalled for breast biopsy (final BI-RADS assessment of 4 or 5) with no breast cancer diagnosed within 1 year |
| Interval cancer rate (i.e., false-negative rate) | Cancer diagnosed within 365 days of a negative study entry screening mammogram | Number of screening exams with a negative assessment (final BI-RADS assessment of 1 or 2) and breast cancer diagnosed within 1 year, per 1,000 screening exams |
| Efficiency metrics (only for the UCLA site) | Through study completion, an average of 1 year | Interpretation time required for radiologists to interpret each mammogram with versus without AI. Delivery time, using time stamp data from exam acquisition to delivery of results to patients (aka turnaround time). |
| Trust and confidence in AI | Years 1,2 and Years 4,5 | Trust and confidence in AI gathered from focus group and survey data |
| False positive recall rate | No cancer diagnosed within 365 days of a positive study entry screening mammogram | Proportion of screening exams recalled for additional imaging (final BI-RADS assessment of 1, 2, or 3), with no breast cancer diagnosed within 1 year |
Countries
United States
Contacts
Primary ContactMichelle L'Hommedieu, PhD
Outcome results
None listed