Ductal carcinoma in situ (DCIS) accounts for a substantial proportion of new breast cancer diagnoses, yet many cases never progress to invasive disease. Surgical excision remains standard care, partly because undersampling at core-needle biopsy means some lesions harbour occult invasion that is only detected at surgery. Improving presurgical risk stratification could help tailor management, including identifying candidates for active monitoring and informing surgical planning.
Radiomic analysis of mammographic calcifications offers a non-invasive approach to characterising lesions beyond visual assessment. A cross-national evaluation spanning the United States, United Kingdom and the Netherlands tested whether radiomics can reliably distinguish pure DCIS from DCIS with occult invasive cancer across heterogeneous populations, imaging platforms and practice patterns, providing evidence for generalisability in real-world settings.
Multinational Cohorts and Heterogeneous Imaging
The analysis drew on digital mammograms from 1498 women diagnosed with DCIS at core-needle biopsy between 2000 and 2021, comprising 696 from the United States, 618 from the United Kingdom and 184 from the Netherlands. Only asymptomatic women with calcifications and without associated masses, architectural distortions or asymmetries were included, with one lesion analysed per patient. Differences in age and lesion size were observed across datasets: the mean age was lower in the Netherlands than in the United States and the United Kingdom, and mean lesion size was smaller in the United Kingdom than in the other two cohorts. These variations reflect clinical diversity that can complicate model transportability.
Must Read: MammosighTR: Nationwide BI-RADS-Labelled Mammograms
Imaging equipment and protocols also differed. The United States cohort included magnification views from GE HealthCare or Hologic systems, the United Kingdom cohort used full-field digital mammography from Hologic, the Netherlands cohort relied on screening full-field digital mammography. Biopsy techniques varied, with vacuum-assisted approaches of differing gauges predominating, and occasional 14-gauge biopsies when calcifications were visible on ultrasound in the United Kingdom and the Netherlands. All samples underwent specimen radiography. Calcifications within radiologist-annotated regions were automatically segmented using a consistent pipeline, and 109 radiomic features capturing shape, morphology and texture at calcification and cluster level were extracted. Features were normalised by vendor and country to mitigate technical variability, and models were trained using repeated cross-validated logistic regression with stability-based feature selection.
Predictive Performance Across Countries
Internal cross-validation areas under the receiver operating characteristic curve (AUCs) were 0.675 for the United States dataset, 0.603 for the United Kingdom and 0.701 for the Netherlands. When the United States-trained model was tested cross-nationally, AUCs were 0.604 in the United Kingdom and 0.682 in the Netherlands, indicating preserved discrimination across settings with differing demographics, imaging and workflows. Pairwise round-robin validation showed a consistent pattern: performance tended to be lower when testing against the United Kingdom dataset, whereas models trained on United Kingdom data performed comparatively better when applied to the other countries. This trend aligns with observed lesion size distributions in the United Kingdom, where smaller upstaged lesions reduced separability from pure DCIS.
Feature selection remained broadly consistent across countries despite site-level differences, supporting method robustness. The most informative features included intensity and spatial descriptors of individual calcifications and clusters, alongside measures of calcification number and geometry. The coherence of selected features across datasets strengthens confidence that the signal exploited by radiomics is not confined to a single imaging protocol or population.
Clinical Scenarios: Active Monitoring and Surgical Planning
To illustrate potential clinical use, operating thresholds were fixed using the United States dataset to target two common scenarios. For active monitoring, a high-sensitivity, high negative predictive value (NPV) threshold was chosen. Applied cross-nationally, NPVs were 89% in the United Kingdom and 92% in the Netherlands, with odds ratios between 1.9 and 2.3, supporting the model’s capacity to help exclude invasion where conservative management is considered. For surgical planning, a high-specificity threshold was selected to prioritise positive predictive value (PPV). Across external tests, PPVs were approximately 30% with odds ratios ranging from 2.4 to 3.2, suggesting value in flagging patients at higher risk of upstaging who may merit discussions about axillary assessment.
Sensitivity and specificity at these fixed thresholds showed expected trade-offs across datasets, yet the operating points clustered tightly by country, indicating concordant behaviour despite heterogeneity. These findings complement earlier work showing that models based solely on clinical features underperform, and they highlight how radiomics may add decision-support value without additional imaging beyond standard mammography.
Radiomic characterisation of mammographic calcifications demonstrated consistent ability to differentiate pure DCIS from lesions with occult invasion across three national datasets, despite variations in patient characteristics, imaging vendors and biopsy practices. Using fixed thresholds tailored to common decision points, the approach delivered high NPV for active monitoring and improved PPV for surgical planning in external validation, indicating practical relevance for pathway selection. Limitations include inter-country differences in screening and biopsy methods, vendor representation and incomplete biomarker data in some cohorts, which warrant cautious interpretation and further investigation. Even so, the cross-national performance and stable feature importance profiles support radiomics as a promising non-invasive adjunct to refine presurgical management in DCIS.
Source: Radiology
Image Credit: iStock
