Novel adversarial semantic structure deep learning for MRI-guided attenuation correction in brain PET/MRI

Abstract

Objective Quantitative PET/MR imaging is challenged by the accuracy of synthetic CT (sCT) generation from MR images. Deep learning-based algorithms have recently gained momentum for a number of medical image analysis applications. In this work, a novel sCT generation algorithm based on deep learning adversarial semantic structure (DL-AdvSS) is proposed for MRI-guided attenuation correction in brain PET/MRI. Materials and methods The proposed DL-AdvSS algorithm exploits the ASS learning framework to constrain the synthetic CT generation process to comply with the extracted structural features from CT images. The proposed technique was evaluated through comparison to an atlas-based sCT generation method (Atlas), previously developed for MRI-only or PET/MRI-guided radiation planning. Moreover, the commercial segmentation-based approach (Segm) implemented on the Philips TF PET/MRI system was included in the evaluation. Clinical brain studies of 40 patients who underwent PET/CT and MR imaging were used for the evaluation of the proposed method under a twofold cross validation scheme. Results The accuracy of cortical bone extraction and CT value estimation were investigated for the three different methods. Atlas and DL-AdvSS exhibited similar cortical bone extraction accuracy resulting in a Dice coefficient of 0.78 ± 0.07 and 0.77 ± 0.07, respectively. Likewise, DL-AdvSS and Atlas techniques performed similarly in terms of CT value estimation in the cortical bone region where a mean error (ME) of less than −11 HU was obtained. The Segm approach led to a ME of −1025 HU. Furthermore, the quantitative analysis of corresponding PET images using the three approaches assuming the CT-based attenuation corrected PET (PET CTAC) as reference demonstrated comparative performance of DL-AdvSS and Atlas techniques with a mean standardized uptake value (SUV) bias less than 4% in 63 brain regions. In addition, less that 2% SUV bias was observed in the cortical bone when using Atlas and DL-AdvSS approaches. However, Segm resulted in 14.7 ± 8.9% SUV underestimation in the cortical bone. Conclusion The proposed DL-AdvSS approach demonstrated competitive performance with respect to the state-of-the-art atlasbased technique achieving clinically tolerable errors, thus outperforming the commercial segmentation approach used in the clinic.

Patch‐based generative adversarial neural network models for head and neck MR‐only planning

“…With deep learning methods, a convolutional neural network consisting of several convolutional layers is trained with data from the two modalities prior to sCT generation. While training is computationally intensive and may require several days, once trained, these methods are extremely fast and produce sCT images in seconds to tens of seconds . Another attractive feature of deep learning techniques is that they can directly extract the relevant set of features from the data without requiring extensive feature engineering.…”

“…While training is computationally intensive and may require several days, once trained, these methods are extremely fast and produce sCT images in seconds to tens of seconds. [13][14][15][16][17][18] Another attractive feature of deep learning techniques is that they can directly extract the relevant set of features from the data without requiring extensive feature engineering. As a result, recent papers have shown promising results from deep learning applied to sCT generation in brain [13][14][15][16][17][18] and pelvis 17,[19][20][21][22][23]…”

Deep-JASC: joint attenuation and scatter correction in whole-body 18F-FDG PET using a deep residual network

“…Pioneering efforts have successfully applied deep convolutional networks to various tasks, including image denoising [13], resolution recovery [14] and image reconstruction [15]. Multiple studies explored the suitability of deep learning approaches for crossmodality transformation from MRI to CT images [16] and vice-versa [17], as well as in AC of PET data [18][19][20]. Recent works focused on the generation of pseudo-CT images from T1-weighted [16,21,22], ultra-short echo time (UTE) [23], zero echo time (ZTE) [24] and Dixon [25] MR sequences for AC of 18 F-FDG PET images in the brain and pelvic regions.…”

MR Image-Based Attenuation Correction of Brain PET Imaging: Review of Literature on Machine Learning Approaches for Segmentation

“…The quality of PET reconstruction is highly dependent on the registration algorithms accuracy. [52] L e v e ls e t S T E [53] L e v e ls e t U T E [55] Thresholding UTE [57] Thresholding ZTE [58] Thresholding Dixon [59] Radon transform T1 weighted [62] Clustering STE and Dixon [63] Clustering T1 weighted [64] Classification DCE, MP-RAGE, T1 weighted [73] Classification Dixon [65] Deep learning T1 weighted [66] Deep learning UTE and out-of-phase echo [84] Deep learning T1 weighted Different atlas-based techniques were proposed in the literature [46,[68][69][70]. Most of the atlas-based methods use machine learning to estimate the pseudo CT image using MR image features such as signal intensity and geometric metrics to learn the relationship between MR signal and Hounsfield units in CT.…”

“…Arabi et al [84] proposed a deep learning generative adversarial semantic model that generates pseudo CT images for MR image-based attenuation correction. The generative adversarial network consists of two main components: synthesis network and segmentation network.…”