Introduction

Dynamic contrast-enhanced (DCE) magnetic resonance imaging (MRI) has been validated for predicting grades of astrocytomas, and differentiating pseudoprogression from true progression¹. More specifically, pharmacokinetic (PK) parameters such as volume transfer constant (K_trans), volume of extravascular extracellular space (Ve), and volume of vascular plasma space (V_p), derived from DCE-MRI can evaluate not only tumor angiogenesis but also permeability of microcirculation. However, to derive the PK parameters robustly, a reliable arterial input function (AIF) that reflects the robust dynamic contrast concentration of blood plasma is crucial, because deriving the PK parameters from DCE-MRI is based on the dynamic distribution of intravenously injected gadolinium-based contrast-agent among tissue compartments, which is called PK modeling. However, DCE-MRI is known for the low reliability², and most of which is caused by irreproducible arterial input function (AIF) measured from the DCE-MRI, or AIF_DCE, due to the relatively low signal intensity of T1-contrast enhanced MR imaging compared to T2* signal intensity of dynamic susceptibility contrast (DSC) MR imaging, another perfusion MRI technique. This results in an AIF obtained from DCE-MRI, or AIF_DCE, sensitive to noise, which leads to PK parameters with low reliability, despite recent advancements. However, dual acquisition of DSC and DCE-MRI requires higher-than-recommended dose of Gadolinium-based contrast agents, which can lead to brain deposition. The aim of this study was to generate more accurate and reliable arterial input function (AIF) using deep learning for DCE-MRI for astrocytomas without additional contrast agents.

Methods

Three hundred eighty-six patients with histopathologically diagnosed astrocytomas who underwent both dynamic susceptibility contrast (DSC) and DCE-MRI preoperatively were enrolled in the study. The AIF was manually obtained from both sequences, AIF_DSC and AIF_DCE (twice for each to examine reproducibility), respectively. A pix2pix model³, one of conditional generative adversarial network (GAN), was developed with the training set (n=260) using AIF_DCE as the input and AIF_DSC as the target to generate a synthetic AIF_{generated DSC}. Using AIF_DCE, AIF_DSC and AIF_{generated DSC}, pharmacokinetic (PK) parameter maps – K_trans, V_e, and V_p – were obtained from the tumor areas in the DCE-MRI images in the test set (n=126). To construct a generator, and a discriminator using convolutional neural network, time-course AIF signal was transformed into a spectrogram image using short-term Fourier transformation. The generator network, a convolutional encoder-decoder, denoise and transformed input images to target images, whereas the discriminator network, fully convolutional network, discriminates the generated images as real or fake (Fig.1). Data was augmented 8 folds, because the AIF was acquired twice for reliability analysis, adding white Gaussian noise, and flipping the spectrogram backward along the time axis. To stabilize training of the GAN model, conditional Wasserstein distance with gradient penalty⁴ was used as loss function. The diagnostic performance in differentiating high-grade from low-grade astrocytomas, obtained by receiver operating characteristic (ROC) analysis, and the reliability of the PK parameters, evaluated by intraclass correlation coefficient (ICC), were compared using the three different AIFs.

Results

The AIF_{generated DSC} was not different to corresponding AIF_DSC, qualitatively (Fig.2). The AIF_{generated DSC}-derived mean K_trans and V_e significantly more accurately differentiated high-grade from low-grade astrocytomas than those derived from AIF_DCE: the area under the ROC curve (AUC) for K_trans, 0.879 vs 0.715, p = 0.0422; V_e, 0.866 vs 0.696, p = 0.0488, respectively (Fig.3). K_trans and V_e showed higher ICCs for AIF_{generated DSC} than for AIF_DCE: K_trans, 0.912 vs 0.383; V_e, 0.682 vs 0.655, respectively (Table 1).

Conclusions

A neural network model can generate an accurate and reliable AIF_{generated DSC} from DCE-MRI to obtain robust PK parameters for differentiating grades of astrocytoma.

Acknowledgements

National Research Foundation of Korea (NRF-2016M3C7A1914448 to K.S.C., NRF-2017M3C7A1031331 to B.J., and NRF-2019K1A3A1A77079379 to S.H.C.); Creative-Pioneering Researchers Program through Seoul National University (SNU) to S.H.C; and Project Code (IBS-R006-D1) to S.H.C.