Acquisition of K^trans perfusion parameter maps from DCE-MRI in breast cancer using a deep learning approach

Traditional methods for calculating perfusion parameter maps such as K^trans require significant computational resources and time, leading to errors due to the variety of analysis models and the difficulty in estimating the arterial input function (AIF). Thus, it is difficult to apply K^trans measurements to clinical diagnosis. The purpose of this study was to investigate whether deep learning (DL) techniques can synthesize K^trans perfusion parameter maps from contrast-enhanced magnetic resonance (MR) images.

A pix2pix-based conditional generative adversarial networks (cGAN) architecture was proposed to generate breast K^trans perfusion maps. The peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) were used to evaluate the quality of the synthetic K^trans maps. Two experienced radiologists were tasked with distinguishing between real and synthesized K^trans maps. The K^trans values of the tumor regions in the synthetic and real K^trans maps were subjected to Pearson correlation analysis and Bland-Altman analysis.

The best performance was obtained when synthesizing K^trans maps using the pix2pix model with spectral normalization (SN) and a local discriminator (LD) (PSNR of 15.167±0.125 and SSIM of 0.690±0.014 for synthesizing K^trans maps from contrast-enhanced MR images). The K^trans values of the tumor regions in the synthetic and real K^trans maps showed a strong correlation (r=0.82), allowed for significant differentiation between benign and malignant tumors (P<0.001), and were not reliably distinguished from real maps by radiologists (accuracy: 41.18%).

The synthesis of breast K^trans perfusion parameter maps from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is enabled by the proposed DL-based method. This provides a new feasible solution for the generation of K^trans.