Breast Cancer
Conditions
Keywords
Mammography Screening, Artificial Intelligence
Brief summary
The purpose of this randomized controlled trial is to assess whether AI can improve the efficacy of mammography screening, by adapting single and double reading based on AI derived cancer-risk scores and to use AI as a decision support in the screen reading, compared with conventional mammography screening (double reading without AI).
Detailed description
European guidelines recommend that mammography exams in breast cancer screening are read by two breast radiologists to ensure a high sensitivity. Double reading is, however, resource demanding and still results in missed cancers. Computer-aided detection based on AI has been shown to have similar accuracy as an average breast radiologist. AI can be used as decision support by highlighting suspicious findings in the image as well as a means to triage screen exams according to risk of malignancy. Eligible women will be randomized (1:1) to the intervention (AI-integrated mammography screening) or control arm (conventional mammography screening). In the intervention arm, exams will be analysed with AI and triaged into two groups based on risk of malignancy. Low risk exams will be single read and high risk exams will be double read. The high risk group will contain appx. 10% of the screening population. Within the high-risk group, exams with the highest 1% risk will by default be recalled by the readers with the exception of obvious false positives. AI risk scores and Computer-Aided Detection (CAD)-marks of suspicious calcifications and masses are provided to the reader(s). In the control arm, screen exams are double read without AI (standard of care). Considering the interplay of number of interval cancers and workload, the study will be considered successful if the interval-cancer rate in the intervention arm is not more than 20% larger than in the control arm. If the interval-cancer rate is statistically and clinically significantly lower in the intervention arm than in the control arm, AI-integrated mammography screening will be considered superior to conventional mammography screening.
Interventions
OTHERAI screening modality
Screen exam will be analysed with an AI system (Transpara, ScreenPoint, Nijmegen, The Netherlands) that assigns exams with a cancer-risk score from 1 to 10, as well as presenting CAD-marks at suspicious findings. Exams with risk score 1-9 will be single read and exam with score 10 will be double read. Risk scores and CAD-marks are provided to the reader(s). The reader(s) will decide whether to recall the woman for work-up or not (as per standard of care). In addition, exams with the highest 1% risk will by default be recalled with the exception of obvious false positives.
OTHERConventional screening modality
Screen exams will be read by two radiologists without the support of AI.
Sponsors
Region Skane
Unilabs
Norwegian Institute of Public Health
Study design
Allocation
RANDOMIZED
Intervention model
PARALLEL
Primary purpose
SCREENING
Masking
SINGLE (Subject)
Masking description
Participants have the possibility to opt-out. If they do not opt-out, neither the participant nor the nurse performing the screen exam will know to what study arm the participant was allocated. The radiologist reading the screen exam will however not be blinded to allocation information.
Eligibility
Sex/Gender
FEMALE
Age
40 Years to 74 Years
Healthy volunteers
Yes
Inclusion criteria
Women eligible for population-based mammography screening.
Exclusion criteria
None.
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Interval-cancer rate | 43 months | Women with interval cancer per 1000 screens |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Cancer-detection rate | 15 months | Women with screen-detected cancer per 1000 screens |
| Recall rate | 15 months | Number of recalls per 1000 screens |
| False-positive rate | 15 months | Women with false positive per 1000 screens |
| Positive Predictive Value-1 | 15 months | Women with cancer for all recalls |
| Sensitivity and specificity | 43 months | True and false-positive rate |
| Cancer detection per cancer type | 19 months | Screen detection of cancer in relation to cancer type, size and stage |
| Tumour biology of interval cancers | 43 months | Characterization of interval cancers per type, size and stage |
| Screen-reading workload | 19 months | Number of screen-readings and number of consensus meetings |
| Incremental cost-effectiveness ratio | 43 months | The incremental cost-effectiveness ratio for AI-integrated mammography screening versus standard of care |
Countries
Sweden
Contacts
PRINCIPAL_INVESTIGATORKristina Lång, MD PhD
Region Skåne
Outcome results
None listed