A Real-Time Centralized Pipeline for Reconstructing and Quantifying Hyperpolarized 129Xe Gas Exchange MRI
Ziyi Wang1, Mu He2, Alexander Culbert1, John Nouls3, Elianna A Bier1, and Bastiaan Driehuys1,3

1Biomedical Engineering, Duke University, Durham, NC, United States, 2Electrical and Computer Engineering, Duke University, Durham, NC, United States, 3Radiology, Duke University Medical Center, Durham, NC, United States


Hyperpolarized 129Xe MRI is emerging as a powerful means to provide 3D quantitative mapping of ventilation, interstitial barrier uptake, and red blood cell transfer. However, this capability requires non-standard radial reconstruction and accurate lung segmentation to enable quantitative analysis. Such reconstruction and image processing would ideally be standardized and centralized to facilitate using 129Xe gas exchange MRI in multi-center clinical trials. To this end, we developed a neural-network based lung segmentation approach that automatically generates accurate masks. With this capability, we demonstrate a fully centralized processing pipeline for real-time reconstruction and quantitative reporting of 129Xe gas exchange MRI.


Recently it has become possible for hyperpolarized 129Xe MRI to provide 3D quantitative mapping of ventilation, interstitial barrier uptake, and red blood cell (RBC) transfer from two inhalations of xenon [1]. However, this capability requires non-standard radial acquisition and reconstruction as well as accurate thoracic cavity segmentation to enable quantitative analysis. To facilitate and accelerate adoption of this capability for multi-center clinical trials, it is desirable to centralize and standardize the reconstruction and image processing procedure. Arguably, the most significant obstacle for such quantitative mapping lies in obtaining an accurate thoracic cavity segmentation. To overcome this hurdle in automatic processing, we have developed an automatic segmentation method based on neural networks, utilizing multi-channel information from both proton and 129Xe images. With this capability, we demonstrate a centralized pipeline for real-time image reconstruction and quantitative reporting of 129Xe gas exchange images.


The study enrolled 80 subjects who underwent an imaging protocol consisting of a calibration spectroscopy, a breath-hold 3D radial 129Xe ventilation/dissolved-phase image, and a breath-hold 1H UTE image, as detailed in [2]. All imaging was conducted on a 3T scanner (Siemens MAGNETOM Trio). The pipeline is designed to facilitate the study without human intervention (Figure 1). It consists of 3 major units to facilitate calibration and parameter setting, image reconstruction and quantification, and clinical reporting (Figure 2). The pipeline was implemented on a workstation running the Ubuntu operating system and operates as follows. First, the 129Xe calibration raw file acquired on the scanner is saved to a shared-access folder, where it is detected by the pipeline’s scout program, which launches its Calibration unit. This calculates parameters for setting scanner frequency, transmit voltages, and the TE required for Dixon-based separation of RBC and barrier images. These results are sent to the MR technologist using the Twilio (https://www.twilio.com) SMS text service, thereby allowing the subsequent imaging sequence to be prescribed. Once the 129Xe gas exchange and anatomical 1H raw data have been acquired and saved to the shared folder, the Image Quantification unit is launched and completes image reconstruction (using compiled C code for speed). After reconstruction, it segments the thoracic cavity (using a neural network model described below) and registers the resulting mask to the functional 129Xe MRI scan (using Simple-ITK). The 129Xe ventilation data is bias-field corrected (N4BiasfieldCorrection[3]), and then thresholded to produce color binning maps of ventilation[4]. Similar processing is applied to barrier uptake and RBC transfer, but since these are expressed as a ratio relative to the gas-phase image, the source images are not bias-field corrected [5]. The maps and the related quantitative metrics are then passed to the Report Generation unit to render clinical report using a HTML template. The report is immediately delivered via an email in PDF and PPT format, making it accessible from mobile devices. The unit also archives NIFTI files of the color maps and populates a running CSV file with the subject ID and the derived quantitative image metrics.

The neural network model for segmentation (Figure 3) employs both the 1H and 129Xe images. It consists of a down-sampling encoder (to extract image features), feature fusion layers (to incorporate the features), and an up-sampling decoder (to restore image resolution), with down-sampling indices preserved to retain high-frequency information [6]. The model was trained using 3D 1H anatomical and 129Xe ventilation images from 116 scans, that were manually segmented in ITK-Snap (http://www.itksnap.org). Training involved minimizing the Mean Squared Error (MSE) loss between the model output and the manually segmented masks.


The calibration result (text message) is delivered around 30 seconds after the calibration scan has been saved. The final clinical report (email) is delivered within 3 minutes of scan completion. Figure 4 shows representative segmentation of the thoracic 1H images using the neural network model. Its accuracy was found to be improved by incorporating 129Xe ventilation image data to clarify regions of unclear boundaries such as the ribs and stomach. Such ventilation imaging was found to be useful even when ventilation defects are present. A representative quantitative report is shown in Figure 5, depicting a COPD patient with ventilation defects, diminished barrier uptake (emphysema) and poor RBC transfer.


The pipeline descried now provides routine, real-time data processing and clinical report generation without human intervention. Moreover, its centralized implementation and rapid delivery positions the system well to ultimately provide remote service through secure, HIPAA-compliant online data transfer. This capability could provide a useful means to standardize image reconstruction and processing procedures to facilitate multi-center studies incorporating gas exchange MRI.


IH/NHLBI R01 HL105643, NIH/NHLBI R01HL126771, and HHSN268201700001C


1. Kaushik, S.S., et al., Single-breath clinical imaging of hyperpolarized (129)Xe in the airspaces, barrier, and red blood cells using an interleaved 3D radial 1-point Dixon acquisition. Magn Reson Med, 2016. 75(4): p. 1434-43.

2. Wang, Z., et al., Hyperpolarized (129) Xe gas transfer MRI: the transition from 1.5T to 3T. Magn Reson Med, 2018.

3. Tustison, N.J., et al., N4ITK: Improved N3 Bias Correction. Ieee Transactions on Medical Imaging, 2010. 29(6): p. 1310-1320.

4. He, M., et al., Extending semiautomatic ventilation defect analysis for hyperpolarized (129)Xe ventilation MRI. Acad Radiol, 2014. 21(12): p. 1530-41.

5. Wang, Z., et al., Quantitative analysis of hyperpolarized 129 Xe gas transfer MRI. Med Phys, 2017. 44(6): p. 2415-2428.

6. Badrinarayanan, V., A. Kendall, and R. Cipolla, SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. Ieee Transactions on Pattern Analysis and Machine Intelligence, 2017. 39(12): p. 2481-2495.

7. He, K.M., et al., Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. 2015 Ieee International Conference on Computer Vision (Iccv), 2015: p. 1026-1034.


The 129Xe central processing server facilitates the acquisition and processing of 129Xe gas exchange MRI. The server reports calibration results within 30 seconds of scan completion/file save and delivers the final clinical report within 3 minutes of completion/saving of the imaging scans. Beyond the technologist operating the scanner and saving data, no additional intervention is required.

The pipeline consists of 3 units: Spectroscopy Calibration, Image Quantification, and Report Generation. The Spectroscopy Calibration unit calculates scanner reference voltage, echo time when RBC and barrier are 90-degree out-of-phase (TE90), and 129Xe gas-phase resonance frequency. These are immediately reported (with SMS text) and used to prescribe the subsequent imaging sequences. The Image Quantification unit fetches files from both the anatomical 1H and functional 129Xe scans, reconstructs and parses the images, and conducts the quantification. The Report Generation unit then produces a clinical report through HTML rendering with the images. The report is delivered within 3 minutes after completion of both scans.

Neural network model inputs anatomic 1H and functional 129Xe images to obtain a thoracic cavity segmentation. The model consists of a down-sampling encoder, feature fusion layers, and an up-sampling decoder. The encoder is composed of convolutional layers (3×3 filter size, number of filters noted on the figure), batch-normalization, PReLU activation [7], and max-pooling layers. The feature-fusion layers use a fully-connected layer (with 25% Dropout) to merge the 8-channel features. The decoder has a structure that is symmetric to that of the encoder to restore the image resolution. The max-pooling indices are saved for unpooling to retain the high-frequency information.

Representative thoracic cavity masks for 3 subjects generated by manual segmentation versus with the convolutional neural network (CNN) model. The model was trained with manual segmentation generated using the active-contour method in ITK-Snap. The CNN model produces accurate lung segmentations, employing both 1H UTE and 129Xe ventilation images. The 129Xe ventilation image provides additional information that resolves unclear lung boundaries caused by the ribs and stomach (second row). Inclusion of 129Xe ventilation still produces high quality lung segmentation, even when substantial ventilation defects are present (third row).

Representative quantitative clinical report in a COPD patient (FEV1: 66%, FEV1/FVC: 68%, FEF25-75: 22%, DLCO: 39%). The left panel shows the distributions of ventilation, barrier uptake and RBC transfer in the lungs of the patient relative to the dashed lines that reflect the healthy reference distributions. The middle panel contains multi-slice montages of each of the 3 maps. The right panel shows the quantitative metrics derived from each map for the patient and shows how they compare to healthy reference values. This patient shows substantial ventilation defects, barrier uptake defects and diminished RBC transfer.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)