Dixon-VIBE Deep Learning (DIVIDE) Pseudo-CT Synthesis for Pelvis PET/MR Attenuation Correction

Abstract

Whole-body attenuation correction (AC) is still challenging in combined PET/MR scanners. We describe Dixon-VIBE Deep Learning (DIVIDE), a deep learning network architecture that allows synthesizing pelvis pseudo-CT maps based only on the standard Dixon volumetric interpolated breath-hold examination (Dixon-VIBE) images currently acquired for AC in commercial Siemens scanners. We propose a network that performs a mapping between the four 2D Dixon MRI images (water, fat, in- and out-of-phase) and their corresponding 2D CT image. In contrast to previous methods, we used transposed convolutions to learn the up-sampling parameters, whole 2D slices to provide context information and pretrained the network with brain images. 28 datasets obtained from 19 patients who underwent PET/CT and PET/MR examinations were used to evaluate the proposed method. We assessed the accuracy of the µ-maps and reconstructed PET images by performing voxel- and region-based analysis comparing the standardize uptake values (SUVs, in g/mL) obtained after AC using the Dixon-VIBE (PETDixon), DIVIDE (PETDIVIDE) and CT-based (PETCT) methods. Additionally, the bias in quantification was estimated in synthetic lesions defined in the prostate, rectum, pelvis and spine. Absolute mean relative change (RC) values relative to CT AC were lower than 2% on average for the DIVIDE method in every region of interest (ROI) except for bone tissue where it was lower than 4% and 6.75 times smaller than the RC of the Dixon method. There was an excellent voxel-by-voxel correlation between PETCT and PETDIVIDE (R2=0.9998, p<0.01). The Bland-Altman plot between PETCT and PETDIVIDE showed that the average of the differences and the variability were lower (mean PETCT-PETDIVIDE SUV=0.0003, σ PETCT-PETDIVIDE=0.0094, CI0.95=[-0.0180,0.0188]) than the average of differences between PETCT and PETDixon (mean PETCT-PETDixon SUV=0.0006, σ PETCT-PETDixo = 0.0264, CI0.95=[-0.0510,0.0524]). Statistically significant changes in PET data quantification were observed between the two methods in the synthetic lesions with the largest improvement in femur and spine lesions. The DIVIDE method can accurately synthesize a pelvis pseudo-CT from standard Dixon-VIBE images, allowing for accurate AC in combined PET/MR scanners. Additionally, our implementation allows rapid pseudo-CT synthesis, making it suitable for routine applications and, even allowing the retrospective processing of Dixon-VIBE data.

Magnetic Resonance-Based Attenuation Correction and Scatter Correction in Neurological Positron Emission Tomography/Magnetic Resonance Imaging—Current Status With Emerging Applications

“…Recently, Liu et al [107] proposed to use only uncorrected PET images with a deep convolutional encoder-decoder network to generate a pseudo-CT. Very recently, a multiparametric MRI model was also suggested to generate pseudo-CT maps based only on Dixon MRI images and was evaluated on head and pelvic images [108]. Among other benefits, these methods also show a great potential for whole-body applications.…”

An overview of deep learning in medical imaging focusing on MRI

“…Kitchen & Seah [216] used GANs to synthetize realistic prostate lesions in T 2 , ADC, K trans resembling the SPIE-AAPM-NCI ProstateX Challenge 2016 38 training data. Other applications are unsupervised synthesis of T1-weighted brain MRI using a GAN [180]; image synthesis with context-aware GANs [217]; synthesis of patient-specific transmission image for PET attenuation correction in PET/MR imaging of the brain using a CNN [218]; pseudo-CT synthesis for pelvis PET/MR attenuation correction using a Dixon-VIBE Deep Learning (DIVIDE) network [219]; image synthesis with GANs for tissue recognition [220]; synthetic data augmentation using a GAN for improved liver lesion classification [221]; and deep MR to CT synthesis using unpaired data [222].…”

Comparison of CBCT‐based dose calculation methods in head and neck cancer radiotherapy: from Hounsfield unit to density calibration curve to deep learning

“…[19][20][21][22][23][24][25] DLMs have been primarily proposed for pCT generation from magnetic resonance imaging (MRI). [26][27][28][29][30][31] They are particularly appealing owing to their fast computation time. One of the first DLMs for pCT generation was based on the U-Net architecture.…”