Advancing breast cancer diagnosis: a combined approach using deep learning-reconstructed diffusion-weighted imaging and Synthetic MRI.
Authors
Affiliations (3)
Affiliations (3)
- Department of Magnetic Resonance, The First Affiliated Hospital of Zhengzhou University, Zhengzhou, China.
- MR Research China, GE Healthcare, Beijing, China.
- Department of Breast Surgery, The First Affiliated Hospital of Zhengzhou University, Zhengzhou, China.
Abstract
Breast cancer is the most common cause of malignancy in women, excluding skin cancer, leading to increased interest in novel non-invasive imaging techniques that eliminate the need for contrast agents. This study aims to establish a deep learning-reconstructed diffusion-weighted imaging (DL-DWI) model combined with Synthetic MRI and to evaluate its diagnostic value for breast cancer in this proof of concept study. A total of 111 patients with pathologically confirmed breast lesions (39 benign, 72 malignant) were enrolled, and all patients underwent DL-DWI and synthetic MRI scans. Quantitative parameters including ADC, DL-ADC, T1, T2 and PD were compared between benign and malignant groups using rank sum tests. ROC curve analysis was used to assess the diagnostic efficacy of single and combined models, and DeLong's test was applied for pairwise AUC comparison. Both DL-ADC (AUC = 0.993; 95% CI: 0.955-1.000) and conventional ADC (AUC = 0.990; 95% CI: 0.949-1.000) demonstrated excellent and comparable diagnostic accuracy. The DL-ADC derived from DL-DWI demonstrated excellent diagnostic accuracy (AUC = 0.993), comparable to conventional ADC (AUC = 0.990). A model combining Synthetic MRI with DL-DWI yielded an AUC of 0.995. However, DeLong's test revealed no statistically significant differences in AUC between any pair of models (all p > 0.05): Synthetic MRI + DL-DWI vs. DL-DWI alone (p=0.45), Synthetic MRI + DL-DWI vs. DWI alone (p=0.54), and DL-DWI vs. DWI alone (p=0.67). In this proof of concept study, standalone DL-DWI achieves high accuracy in differentiating benign and malignant in breast lesions meeting the study size criteria. Adding synthetic MRI brings no significant diagnostic improvement, indicating DL-ADC is the dominant diagnostic biomarker. This contrast-free deep learning-based MRI approach has potential as a contrast free protocol and should be further evaluated in a larger clinical trial comparing contrast-enhanced breast MRI with this novel technique.