ML models of radiomics on MRI predict upstage of pre-cancerous breast lesions to invasive breast cancer on surgery.

• Background: Ductal carcinoma in situ (DCIS) is a non-lethal, non-invasive precancerous breast lesion that can exist with or recur as invasive breast cancer. While roughly a quarter of DCIS cases upstage to invasive disease, a majority of patients diagnosed with DCIS undergo surgery and radiation therapy, up to half of which are over-treated. Thus, there is a need for robust risk stratification strategies that can identify patients at low risk of upstaging that can instead be monitored instead of subjected to aggressive treatments. 
• Methods: In a multicenter trial, we compute radiomic features on MRI and summarize the feature space through hierarchical clustering (i.e., radiomic phenotyping) and principal component (PC) analysis. From there, we build logistic regression models of disease upstaging by combining the top radiomic PCs and radiomic phenotypes with clinical factors. 
• Results: The logistic regression model parameterized by the top 3 radiomic PCs and clinical factors enabled the identification of an additional 25% true negatives (no disease upstaging) compared to models with clinical factors alone. Radiomic features therefore show promise in risk stratification among DCIS patients, which may allow low risk patients to be confidently identified for less aggressive treatment like active surveillance. 

Radiomics on MRI are associated with minimal residual disease in bone marrow aspirate samples in post-treatment patients.