Decision support in breast cancer screening Cancer Malignant neoplasm of breast
Conditions
Interventions
Data will be collected retrospectively from the OPTIMAM Mammography Image Database, with patient consent. There will be no impact on patient care. The intervention is the AI system, assessed on de-ide
Sponsors
Imperial College London
Eligibility
Sex/Gender
Female
Age
50 Years to 70 Years
Inclusion criteria
Inclusion criteria: 1. Women undergoing routine breast cancer screening (age 50-70 years) as part of the national breast screening programme from January 2016 onwards 2. Mammography images acquired using Hologic/Lorad, Siemens, or GE devices
Exclusion criteria
Exclusion criteria: Part A: 1. Women attending an assessment clinic or symptomatic clinic (i.e. not routine screening) 2. Women undergoing annual screening due to: 2.1. High risk (lifetime risk >30% - e.g. faulty BRCA1, BRCA2, TP53) 2.2. Moderate risk (lifetime risk 17-30%) 2.3. Personal stratified follow up (e.g. indeterminate B3 lesions) 3. Presence of breast implants 4. Screens with incomplete (<4 standard screening views - e.g. due to abandoned screen) 5. Poor diagnostic quality imaging (which would be repeated) 6. Non-standard acquisitions beyond the routine 4 screening views 7. For negative or benign cases, women without a negative follow up screen approximately 3 years later (at least 31 months after initial screen), as this would preclude determination of a robust ground truth Part B: Same dataset as defined in Part A, with the same inclusion and exclusion criteria
Design outcomes
Primary
| Measure | Time frame |
|---|---|
| Sensitivity and specificity of AI system cancer detection measured as the number of positive cases (cases considered positive if they received a biopsy-confirmed diagnosis of cancer within 39 months following the screening visit. Negative cases will require a negative result from the study screening visit, and another negative result at the subsequent screening visit at least 31 months later) compared to first, second and consensus reader decisions. | — |
Secondary
| Measure | Time frame |
|---|---|
| 1. Case recall rate, cancer detection, positive predictive value, negative predictive value, cancer detection rate, area under the receiver operating characteristic curve will be measured for AI system performance over the study dataset time period 2. Subgroup performance by factors including cancer type and grade, primary tumour size, patient age, breast density, prior cancer, prevalent and incident screens, ethnicity, device manufacturer, socioeconomic status, and screening site over the study dataset time period 3. Analysis of failure cases for the study dataset time period 4. Percentage of women that meet the eligibility criteria over the course of the study 5. Simulations of workforce impact assessment and health economic modelling over the study period 6. AI system localisation performance (if lesion position data available) over the study period 7. AI system performance in confirmed interval cancers (percentage of historical interval cancers that the AI system flagged for recall, and qualitative agreement of the localisation in the original screening mammogram with the presence/absence of true radiological evidence) over the study period | — |
Countries
England, United Kingdom
Outcome results
None listed