Deep learning reconstruction for accelerating synthetic magnetic resonance imaging of the breast: a comparative...

Synthetic magnetic resonance imaging (SyMRI) offers synchronized acquisition of multiparametric maps and contrast-weighted images, showing promise in breast cancer diagnosis. However, its relatively long scan time limits broader clinical application. The aim of this study was to evaluate the feasibility of deep learning reconstruction (DLR) in accelerating SyMRI and to compare its image quality, quantitative accuracy, and diagnostic performance in breast imaging with those of standard SyMRI.

Fifty-eight female patients (mean age 50.38±1.34 years) with suspected breast cancer underwent three SyMRI protocols: standard 2× SyMRI, accelerated 3× SyMRI, and 3× SyMRI with DLR (3× SyMRI-DLR). Quantitative parameters [T1, T2, and proton density (PD)], signal-to-noise ratio (SNR), and subjective image quality (e.g., overall image quality, anatomical clarity, diagnostic information, tissue contrast, uniformity, and artifacts) were evaluated. Receiver operating characteristic (ROC) analysis and the area under the curve (AUC) were used to evaluate the diagnostic performance in identifying breast lesions and distinguishing triple-negative breast cancer (TNBC) from non-TNBC.

The 3× SyMRI and 3× SyMRI-DLR protocols reduced the scan time compared to 2× SyMRI. No significant differences were observed in T1, T2, or PD values across the three protocols (all P values >0.05), and there were strong correlations between 3× SyMRI-DLR and 2× SyMRI (r_s>0.89; P<0.001). The SNR of 3× SyMRI-DLR was significantly higher than that of 3× SyMRI [T1-weighted imaging (T1WI): P<0.001; T2-weighted imaging (T2WI): P<0.001] but not significantly different from that of 2× SyMRI for T1-weighted imaging (P=0.343). Subjective evaluation showed that 3× SyMRI-DLR had superior anatomical clarity and image uniformity to 3× SyMRI (all P values <0.001), although its artifacts were more prominent than those of 2× SyMRI. Quantitative T1 and T2 values were significantly higher in malignant lesions than in normal breast tissue in all protocols (P<0.05). Diagnostic performance for breast cancer detection using T1 and T2 values was similar among the three sequences, with AUCs for ranging from 0.857 to 0.858 for T1 and from 0.625 to 0.632 for T2. In distinguishing TNBC from non-TNBC, T2 values yielded an AUC of 0.721 [95% confidence interval (CI): 0.587-0.832] in the 3× SyMRI-DLR group (P<0.05), which was similar to that of the standard sequences.

Integrating DLR into accelerated SyMRI for breast imaging yielded a 37.46% reduction in scan time compared to that of the standard SyMRI while maintaining quantitative accuracy and image quality.