Breast Neoplasm Female
Conditions
Keywords
Artificial Intelligence, Breast Cancer, Mammography, Screening
Brief summary
This is a prospective clinical trial following a paired screen-positive design, with the aims to assess the performance of an artificial intelligence (AI) computer-aided detection (CAD) algorithm as an independent reader, in addition to two radiologists, of screening mammograms in a true screening population. Since all decisions by individual readers will be recorded, it is possible to determine what the outcome would have been had one or two of the readers not been allowed to assess images, and to determine what the outcome would have been had the recall decision been performed by consensus decision (actual) compared to single reader arbitration of discordant cases.
Interventions
DIAGNOSTIC_TESTAI CAD
The Lunit INSIGHT MMG will be used as the AI CAD in our study. Initially, version 1.6.1.1 will be installed. The software version will be continuously updated with subsequent software releases, after confirming in a historic calibration dataset that the performance is improved. The operating point will be set based on a historic calibration dataset to attain a joint sensitivity of breast cancer detection of AI and first reader which is 2% higher than for first and second reader.
DIAGNOSTIC_TESTRadiologist reading
Standard of care, each radiologist will assess the mammography examination, making a binary flagging decision (flag the examination to continue to consensus discussion, or not)
Sponsors
Capio Sankt Görans Hospital
Lunit Inc.
Karolinska Institutet
Karolinska University Hospital
Study design
Allocation
NON_RANDOMIZED
Intervention model
SINGLE_GROUP
Primary purpose
DIAGNOSTIC
Masking
TRIPLE (Subject, Caregiver, Investigator)
Masking description
Positive disease status is ascertained by pathology-verified breast cancer. Disease status is not known to any of the actors (except for the outcomes assessor by necessity). AI decision is not known by the care provider radiologists until they have made their decisions. In the subsequent consensus discussion where a decision is made to recall or not to recall a woman, the AI decision is known. After AI decision has been recorded and outcomes have been assessed, the investigators will have full information on outcomes and AI decisions.
Intervention model description
This is a prospective clinical trial following a paired screen-positive design (Pepe, Alonzo; 2001), with the aims to assess the performance of an AI algorithm combined with radiologists(s) compared to standard-of-care being two radiologists assessing screening mammograms in a true screening population. Since all decisions by individual readers will be recorded, it is possible to determine what the outcome would have been had one or two of the readers not been allowed to assess images, and to determine what the outcome would have been had the recall decision been performed by consensus decision (actual) compared to single reader arbitration of discordant cases.
Eligibility
Sex/Gender
FEMALE
Age
40 Years to 74 Years
Healthy volunteers
No
Inclusion criteria
* Participants in regular population-based breast cancer screening at Capio St Göran Hospital
Exclusion criteria
* Incomplete exam (complete exam: mediolateral oblique and craniocaudal images of Left and Right breast) * Breast implant * Complete mastectomy (excluded from screening positive group) * Participant in surveillance program for prior breast cancer
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Incident breast cancer | At Screening | Breast cancer diagnosis by pathologist |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Reader flagging | At screening | Radiologist or AICAD assessing the mammograms as suspicious or not suspicious for malignancy |
| Consensus recall | At screening | A decision by the consensus discussion to recall the woman for further work-up |
| Tissue sampling | At screening | Biopsy or fine needle aspiration performed |
| Process failure | At screening | Failure of the AI CAD software to generate AI scores |
Countries
Sweden
Outcome results
None listed