Virtual dynamic contrast enhanced breast MRI using 2D U-Net Architectures.

Abstract

Breast Magnetic Resonance Imaging (MRI) examinations routinely include contrast-agent based dynamic contrast-enhanced (DCE) acquisitions. Expanding the accessibility and personalization of breast MRI might be supported amongst others by advancing non-contrast-enhanced MRI, such as virtual dynamic contrast-enhanced techniques (vDCE) utilizing neural networks. This IRB-approved retrospective study includes n=540 breast MRI examinations acquired on a single 3T MRI scanner. Two 2D U-Net architectures were trained using non-contrast-enhanced MRI acquisitions including T1w, T2w and multi-b-value diffusion weighted imaging acquisitions as inputs and either a single (SCO-Net) or multiple (MCO-Net) time points of a DCE series as ground truth. The neural networks predicted a vDCE series corresponding to five consecutive DCE time points. Across all time points, no significant differences in structural similarity index (SSIM) could be found between the SCO-Net and MCO-Net, both achieving a mean SSIM of 0.86. For peak-signal-to-noise-ratio and normalized root-mean-square error, significantly better results could be observed for the MCO-Net reaching scores of 24.42dB and 0.087 respectively. Comparison of manual segmentations of findings on DCE and vDCE images reached a DICE score of 0.61 and an intersection over union (IoU) of 0.47 without significant differences between SCO-Net and MCO-Net. These findings suggest a technical feasibility of generating vDCE image series from unenhanced input acquisitions using neural networks. However, the analysis does not allow drawing any conclusion on the clinical assessment of lesion specific curve kinetics, which need to be assessed prior determining on the feasibility of deriving diagnostically meaningful enhancement characteristics in individual lesions.