Cross-national radiomics validation using mammography to predict occult invasion in ductal carcinoma in situ

Abstract

Background: Patients diagnosed with ductal carcinoma in situ (DCIS) may also have undetected invasive breast cancer. Radiomic features of calcifications at mammography can predict occult invasive disease among women diagnosed with DCIS at core-needle biopsy, which could affect treatment recommendations. However, the generalizability of these radiomic models must be tested before they are adopted in clinical practice. Purpose: To evaluate the generalizability of radiomic models based on mammography features to predict occult invasive cancer among women diagnosed with DCIS at core-needle biopsy from three national datasets. Materials and Methods: In this retrospective, cross-national study, digital mammograms from women diagnosed with DCIS at breast core-needle biopsy were collected in the United States, United Kingdom, and the Netherlands between January 1, 2000, and December 31, 2021. Only asymptomatic women who had calcifications but did not have associated masses, architectural distortions, or asymmetries were included. Radiomic models were developed using cross-validated logistic regression on each national dataset, then round-robin tested on the other datasets. Differences across the three datasets in terms of the upstaging rate, age, lesion size, and estrogen and progesterone receptor levels were assessed using Kruskal-Wallis or χ2 test. Results: The study included 1498 women (age range, 31–89 years; mean age, 59 years ± 9 [SD]), as follows: 696 women from the United States, 618 women from the United Kingdom, and 184 women from the Netherlands, with upstaging rates of 16.1%, 16.7%, and 14.1%, respectively. Internal cross-validation areas under the receiver operating characteristic curve (AUCs) were 0.675 (95% CI: 0.671, 0.679), 0.603 (95% CI: 0.567, 0.722), and 0.701 (95% CI: 0.697, 0.706) for the U.S., UK, and Netherlands datasets, respectively. The model that was trained on the U.S. dataset yielded cross-national validation AUCs of 0.604 (95% CI: 0.560, 0.648) and 0.682 (95% CI: 0.607, 0.757) for the UK and Netherlands datasets. Conclusion: Radiomic machine learning models were shown to have the potential to predict occult invasive cancer in women with DCIS across diverse settings.