Deep Learning-Based Synthetic Contrast-Enhanced Breast MRI for Monitoring Response to Neoadjuvant Therapy

Background/Objectives: Contrast-enhanced (CE) breast MRI is highly sensitive for evaluating breast cancer extent and response to neoadjuvant therapy (NAT) but requires intravenous administration of gadolinium-based contrast agents (GBCA), increasing cost, time, patient discomfort, and health concerns. This study explored the feasibility of reducing GBCA use in treatment monitoring using a deep learning (DL) model to synthesize CE-MRI from non-contrast MRI. Methods: This IRB-approved retrospective pilot study evaluated women with breast cancer enrolled in an ongoing trial using serial MRI to monitor NAT prior to surgery. A pre-trained DL model was used to synthesize CE-MRI from T1-, T2-, and diffusion-weighted MRI. Changes in tumor volume at early (post-1-cycle NAT) and mid-treatment were measured on synthetic and acquired CE-MRI. Performance for predicting residual cancer burden (RCB) class 0/1 was evaluated using AUC and compared with DeLong’s test. Results: 27 women were included in the study (median age, 47 years range = 28–75); 14 (52%) achieved RCB class 0 and six (22%) achieved class 1. Synthetic CE-MRI-derived tumor volumes showed strong correlation with those from acquired CE-MRI at pre-treatment (ρ = 0.92, p < 0.001) and early treatment (ρ = 0.83, p < 0.001), but lower agreement at mid-treatment (ρ = 0.57, p = 0.002). Change in tumor volume on synthetic CE-MRI was numerically similar to acquired CE-MRI for predicting RCB class 0/1 vs. 2/3 at both early (AUC = 0.84 vs. 0.86, p = 0.83) and mid-treatment (AUC = 0.73 vs. 0.75, p = 0.80). Conclusions: Synthetic CE-MRI demonstrates preliminary feasibility as a non-contrast surrogate for predicting favorable outcomes (RCB class 0/1) in this pilot study, but inconsistencies in tumor volume measurement vs. acquired CE-MRI warrant further model refinement and validation.