A novel supervised learning method to generate CT images for attenuation correction in delayed pet scans

https://doi.org/10.1016/j.cmpb.2020.105764Get rights and content

Highlights

  • A reconstruction network is developed to convert PET raw data into pseudo PET image.

  • An elaborate network is designed to generate CT image for attenuation correction in PET image reconstruction. Inputs of the network are the two pseudo PET images in the first and delayed scan, and the CT image in the first scan. The network can output an estimated CT image in the delayed scan.

  • This method only requires one CT scan and therefore significantly reduces the X-ray radiation dose received by the patient in delayed PET imaging.

Abstract

Background and objectives

Attenuation correction is important for PET image reconstruction. In clinical PET/CT scans, the attenuation information is usually obtained by CT. However, additional CT scans for delayed PET imaging may increase the risk of cancer. In this paper, we propose a novel CT generation method for attenuation correction in delayed PET imaging that requires no additional CT scans.

Methods

As only PET raw data is available for the delayed PET scan, routine image registration methods are difficult to use directly. To solve this problem, a reconstruction network is developed to produce pseudo PET images from raw data first. Then a second network is used to generate the CT image through mapping PET/CT images from the first scan to the delayed scan. The inputs of the second network are the two pseudo PET images from the first and delayed scans, and the CT image from the first scan. The labels are taken from the ground truth CT image in the delayed scan. The loss function contains an image similarity term and a regularization term, which reflect the anatomy matching accuracy and the smoothness of the non-rigid deformation field, respectively.

Results

We evaluated the proposed method with simulated and clinical PET/CT datasets. Standard Uptake Value was computed and compared with the gold standard (with coregistered CT for attenuation correction). The results show that the proposed supervised learning method can generate PET images with high quality and quantitative accuracy. For the test cases in our study, the average MAE and RMSE of the proposed supervised learning method were 4.61 and 22.75 respectively, and the average PSNR between the reconstructed PET image and the ground truth PET image was 62.13 dB.

Conclusions

The proposed method is able to generate accurate CT images for attenuation correction in delayed PET scans. Experiments indicate that the proposed method outperforms traditional methods with respect to quantitative PET image accuracy.

Introduction

PET/CT imaging systems are widely used in clinical diagnosis and treatment [1]. CT provides anatomical information to localize and segment lesions and tumors [2]. A PET image, on the other hand, provides metabolic information to further aid in the clinical diagnosis [3]. Because the gamma rays are attenuated when they pass through the body, attenuation correction is critical in PET imaging to quantitatively reconstruct the radiotracer distribution [4,5].

For PET/CT systems, attenuation correction is commonly performed with CT images [6,7]. A CT image is essentially an attenuation map of an X-ray inside the body. Since X-rays and gamma rays are both types of electromagnetic radiation, researchers have established models to convert X-ray attenuation maps into gamma ray attenuation maps. Then a gamma ray attenuation map can be used to reconstruct a PET image. Many methods have been proposed to convert X-ray attenuation maps into the gamma ray attenuation maps. The most widely used approach is the bilinear energy-mapping method [8] that was proposed by Bai et al. [9]. Based on the assumption that all tissues can be described as a mixture of water-air for “soft tissue” and water-cortical bone for “bones”, they demonstrated that there is a functional relation between the CT number H and the linear attenuation coefficient μ.This method has been commonly implemented by manufacturers for attenuation correction in PET/CT systems [10].

In some applications, when CT images are difficult to acquire, researchers have proposed various solutions based primarily on image registration and image synthesis. Recently, deep learning based approaches have become popular. The mainstream non-rigid registration methods based on deep learning are non-supervised learning and supervised learning. A supervised learning method defines an objective function containing the ground truth labels and optimizes the neural network weights based on a backpropagation procedure. Vos et al. [11] proposed an end-to-end unsupervised deformable image registration network named DIRNet. The network consists of a convolutional neural network regressor, a spatial transform network and a resampler, and the B-spline model is used to represent the non-rigid transformation. Balakrishnan et al. [12] developed VoxelMorph, an unsupervised learning-based framework for non-rigid pairwise medical image registration; a U-Net like structure is used to generate the output, and a dense transform field, which contains the motion vector for every pixel, is used to represent the non-rigid transformation. Zhao et al. [13,14] proposed a U-Net like structure named VTN for medical image registration, and a cascaded network strategy was utilized to improve performance. Unlike unsupervised learning method, in a supervised learning method, a ground truth mask or label is provided for each input data set; these methods commonly outperform the unsupervised learning methods [15]. Sokooti et al. [16] proposed a supervised learning framework for non-rigid image registration. The method uses a DVF (Deformable Vector Field) to generate the input image pair and sets the motion field as the ground truth label. Eppenhof et al. [17] presented a non-rigid registration method based on a 3D convolutional neural network. The network was trained on synthetic random transformations, which is applied to a small set of representative images for desired application. Uzunova et al. [18] adapted the FlowNet architecture for CNN-based flow estimation to the non-rigid medical image registration; they proposed a novel approach for developing appearance models from few training samples and synthesized large amounts of realistic ground truth data to train the CNN network.

Since X-ray radiation may increase the risk of cancer, many methods have been proposed to synthesize CT images from images of other modalities for PET attenuation correction. Roy et al. [19] proposed a patch-based method to generate whole head attenuation maps from ultrashort echo-time (UTE) MR imaging sequences and a reference dataset; then a Bayesian framework was used to acquire the corresponding relations. Liu et al. [20] proposed an automated approach to generate discrete-valued pseudo CT scans (for soft tissue, bone, and air) from a single high-spatial-resolution diagnostic-quality three-dimensional MR images. Raw MR images were used as input, and reference tissue labels for air and bone were drawn to train the network. The output CT image was used to implement the PET attenuation correction. The results demonstrated that the method can provide reduced PET reconstruction errors relative to a CT-based standard within the brain compared with those produced by the current MR imaging-based attenuation correction approaches. Spuhler et al. [21] used a CNN to generate patient-specific transmission data from MRI images, and this approach maintained sufficient accuracy for both static and dynamic PET studies. Nie et al. [22] proposed a GAN-based method to generate a CT image from an MR image. The AutoContext model was applied to implement a context-aware generative adversarial network. Wolterink et al. [23] proposed a GAN-based model for CT-to-MR and MR-to-CT conversion, and CycleGAN was used to implement the forward and backward cycles; in this approach, unpaired and unaligned MR and CT images were used to train the model.

In delayed PET image acquisition, CT scans are usually performed twice to provide attenuation correction for accurate delayed PET image reconstruction. Considering that there is no dramatic change in the anatomical structure of a patient between the two acquisitions, the CT image needed for attenuation correction can potentially be obtained via image generation. In this paper, we propose a method to generate CT images for delayed PET attenuation correction. A reconstruction network is first developed to produce pseudo PET images from raw data, and then an elaborate network is designed to generate the CT image for the delayed PET scan. The input of the network consists of the CT and pseudo PET images at T1, and the pseudo PET image at T2. The output is the CT image at T2. This method may reduce the X-ray radiation dose of the patient in delayed PET/CT scans, or more general PET/CT scans.

The remainder of the paper is organized as follows. In Section 2, a novel framework based on supervised learning is presented, and the implementation of the approach is described. In Section 3, the proposed method is evaluated with simulated and clinical datasets. The discussion and conclusion are presented in Sections 4 and 5 respectively.

Section snippets

Method

Fig. 1 shows the workflow of delayed PET imaging using PET/CT, which is the most common scenario in practical applications. For patients undergoing delayed PET/CT scans or multiple scans during therapy, it is generally not practical to have the patients in the same position without motion. Therefore, CT scans have to be performed at multiple acquisition times for PET attenuation correction, which significantly increases exposure to the radiation dose. To avoid repeated CT scans and obtain the

Experiment

We applied the proposed method to generate CT images for PET attenuation correction, and the proposed method was compared with VoxelMorph and ResUNet. CT image generation and PET image reconstruction with attenuation correction experiments were conducted. In CT image generation, the proposed method was compared with other methods, and the quality of the output CT images was assessed. In attenuation correction for PET, a comprehensive procedure for PET image reconstruction was developed using

Discussion

We propose a novel CT generation method for attenuation correction in delayed PET imaging. A reconstruction network is developed to convert PET raw data to pseudo PET images, and then a supervised learning based network is designed to generate CT images for delayed PET scans using previously obtained CT and PET data and the delayed PET data.

The results of the PET attenuation correction experiment indicate that the proposed method performs better than VoxelMorph and ResUNet. The mean εMAE and

Conclusion

A novel method based on supervised learning is proposed for attenuation correction in delayed PET imaging. A reconstruction network is developed to convert PET raw data into pseudo PET images. Then a supervised learning framework is utilized to generate CT images for attenuation correction with the first PET/CT dataset and the pseudo PET image from the delayed scan. This method only requires one CT scan and therefore significantly reduces the X-ray radiation dose received by the patient in

Declaration of Competing Interest

None.

Acknowledgments

This research is supported by the China Postdoctoral Science Foundation (No. 2020M671827) and Zhejiang Lab.

References (26)

  • S. Afshari et al.

    Automatic localization of normal active organs in 3D PET scans

    Comput. Med. Imaging Graph.

    (2018)
  • C.T. Mesina et al.

    Effects of attenuation correction and reconstruction method on PET activation studies

    Neuroimage

    (2003)
  • A. Kaalep et al.

    EANM/EARL FDG-PET/CT accreditation-summary results from the first 200 accredited imaging systems

    Eur. J. Nucl. Med. Mol. Imaging

    (2018)
  • H.R. Shahdoosti et al.

    MRI and PET image fusion using structure tensor and dual ripplet-II transform

    Multimed. Tools Appl.

    (2018)
  • S.K. Øen et al.

    Quantitative and clinical impact of MRI-based attenuation correction methods in [18F] FDG evaluation of dementia

    EJNMMI Res.

    (2019)
  • X. Dong et al.

    Deep learning-based attenuation correction in the absence of structural information for whole-body PET imaging

    Phys. Med. Biol.

    (2020)
  • S.L. Brady et al.

    Dose optimization: a review of CT imaging for PET attenuation correction

    Clin. Transl. Imaging

    (2017)
  • W. Territo et al.

    The role of CT technique and object size on PET image ROI quantitation and variability

    J. Nucl. Med.

    (2017)
  • M.R. Ay et al.

    Comparative assessment of energy-mapping approaches in CT-based attenuation correction for PET

    Mol. Imaging Biol.

    (2011)
  • C. Bai et al.

    A generalized model for the conversion from CT numbers to linear attenuation coefficients

    IEEE Trans. Nucl. Sci.

    (2003)
  • M. Oehmigen et al.

    Improving the CT (140 kVp) to PET (511keV) conversion in PET/MR hardware component attenuation correction

    Med. Phys.

    (2020)
  • B.D. de Vos et al.

    End-to-end unsupervised deformable image registration with a convolutional neural network

    Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support

    (2017)
  • G. Balakrishnan et al.

    VoxelMorph: a learning framework for deformable medical image registration

    IEEE Trans. Med. Imaging

    (2019)
  • Cited by (0)

    View full text