Cross-National Radiomics Validation Using Mammography to Predict Occult Invasion in Ductal Carcinoma in Situ

Area of Science:

• Radiology
• Oncology
• Machine Learning in Medicine

Background:

• Ductal carcinoma in situ (DCIS) diagnosis may mask underlying invasive breast cancer.
• Mammographic calcifications' radiomic features can predict occult invasive disease.
• Generalizability of predictive models requires validation across diverse populations.

Purpose of the Study:

• To assess the generalizability of radiomic models for predicting occult invasive cancer.
• Models are based on mammography features in women diagnosed with DCIS.
• Evaluation performed across three national datasets.

Main Methods:

• Retrospective, cross-national study of digital mammograms from US, UK, and Netherlands.
• Included asymptomatic women with calcifications but no masses or distortions.
• Radiomic models developed using cross-validated logistic regression and round-robin tested.

Main Results:

• 1498 women included; upstaging rates ranged from 14.1% to 16.7%.
• Internal cross-validation AUCs ranged from 0.603 to 0.701.
• Cross-national validation AUCs demonstrated model applicability across datasets.

Conclusions:

• Radiomic machine learning models show promise in predicting occult invasive cancer in DCIS patients.
• Models demonstrated generalizability across diverse national settings.
• These findings support potential clinical utility in treatment recommendations.