Development and internal validation of mammography feature-based prognostic models for distant recurrence-free survival of invasive breast cancer in a screening cohort.

Abstract

Overdiagnosis in breast cancer screening may be reduced by identifying lesions that, although detected on screening mammograms, are unlikely to progress to poor outcomes and may not require recall. As a proof-of-concept, we evaluated prognostic models for 10-year distant recurrence-free survival (DRFS) using radiomics features from invasive breast cancers (IBCs) presenting as masses on screening mammograms. In a cohort of screened women, 1466 IBC patients with clinical, pathological, and follow-up data were identified through national registries. Using radiomics features, tumor volume, and specific growth rate, proportional hazards models were developed to predict 10-year distant recurrence risk. Models were trained using positive screening mammograms of patients with screen-detected IBC (n = 1063) and diagnostic mammograms of patients with interval cancer (n = 406). Performance was evaluated only in screen-detected IBCs using repeated nested cross-validation. Median follow-up was 5.1 years (10th-90th percentile: 2.1-10.1), with 111 distant recurrences within 10 years. Model performance was moderate (C-index 0.70 [SD 0.01], calibration slope 1.22 [SD 0.13]), with predicted 10-year recurrence risks ranging from 4.9% to 18.0% (10th-90th percentile). The 42 IBCs with the lowest predicted risk had a 10-year recurrence probability of 2.6% (95% CI 0-7.5%), of whom only five received adjuvant systemic therapy. These findings suggest that mammography-based radiomics features may help identify low-risk IBCs and potentially reduce overdiagnosis by reconsidering recall for selected cases.