Image Analysis With Artificial Intelligence to Increase Precision in Breast Cancer Screening - the ScreenTrust MRI Substudy: a Prospective Trial of AI to Select Women for Supplemental Screening MRI
Image Analysis With Artificial Intelligence to Increase Precision in Breast Cancer Screening - the ScreenTrust MRI Substudy: a Prospective Trial of AI to Select Women for Supplemental Screening MRI
This is a prospective clinical trial aiming to determine the ability of an AI pipeline to identify women who would benefit from supplemental MRI in terms of decreasing the number of cancers having a significantly delayed detection
All women attending mammography screening at Karolinska University Hospital will have their mammograms analyzed by AI (Figure 1). The specific AI-implementation (AI tool) in this study is a result of AI predictions from three equally weighted component AI models analyzing mammograms: (i) masking predictor, (ii) risk predictor and (iii) cancer signs predictor (by one commercial CAD model and one in-house academic CAD model); the age of the woman is also taken into account by multiplying the score with (110-age)/70. The purpose of the age factor is to attain a relatively similar proportion of MRI exams in the lower and higher age groups. The aim of the AI tool is to identify women with the highest probability of having a delay in cancer detection, i.e., having had a false negative screening mammogram.
An AI-based framework has been developed by researchers at Karolinska Institute (led by Dr. Fredrik Strand) and Royal Institute of Technology (led by Dr: Kevin Smith). The specific AI-implementation (AI tool) in this study is a result of AI predictions from three equally weighted component AI models analyzing mammograms: (i) masking predictor, (ii) risk predictor and (iii) cancer signs predictor (by one commercial CAD model and one in-house academic CAD model); the age of the woman is also taken into account by multiplying the score with (110-age)/70. The purpose of the age factor is to attain a relatively similar proportion of MRI exams in the lower and higher age groups. The aim of the AI tool is to identify women with the highest probability of having a delay in cancer detection, i.e., having had a false negative screening mammogram. The specific AI tool and its settings will remain the same during the study. For each examination, the AI tool will produce an AI Joint Score and an AI Masking Score. The AI Masking Score cut-off point was defined by the median of examinations collected during the initial period of March 1 to March 24, 2021. The cut-off point of the AI Joint Score was defined by the 92nd percentile of the initial population. Women meeting these criteria will be invited to the study, and randomized to MRI or no-MRI (standard-of-care).
A Signa Premier 3T MRI scanner from GE Healthcare will be used. The MRI protocol will contain a T2-weighted Dixon sequence and a T1-weighted dynamic contrast enhanced series, and will remain the same through the course of the study. All MRI exams will be assessed by two radiologists, where the second reader will have access to the assessment of the first reader. In case of disagreement, a consensus discussion between two radiologists will be held. The MRI exams will be assessed according to BI-RADS, and follow-up will depend on the BI-RADS category (Figure 2). Women with BI-RADS 1-2 will have no further diagnostics and will be sent a 'healthy letter'. Women with BI-RADS 3 to 5 will be recalled for 2nd look ultrasound. Women with BI-RADS 4-5 will be included in the regular process for established cancer suspicion and be discussed in a multidisciplinary team conference. For women with BI-RADS 3, the follow-up will be handled within the breast radiology unit.
Inclusion Criteria:
Exclusion Criteria: