Generating CT images in delayed PET scans using a multi-resolution registration convolutional neural network
Introduction
The positron emission tomography/computed tomography (PET/CT) system provides crucial information for radiation treatment, making critical decisions when diagnosing tumour, prognosis and staging [1]. A delayed PET/CT scan refers to scanning the selected bed positions again within a certain period of time after the regular PET/CT scan to make a more accurate diagnosis of the disease [2]. At the same time, precise PET images demand the anatomical information provided by CT images, but this will increase the overall X-ray radiation to the patient [3]. Therefore, finding a method to generate delayed CT (T2CT) to avoid additional CT scans is particularly important.
The development of deep learning (DL) has shown tremendous potential in the field of medical imaging processing and its application [4]. Various DL techniques such as segmentation and reconstruction are being used to build computer-aided diagnosis and treatment systems [5], [6]. Recently, several convolutional neural network (CNN) models have shown great results in the field of image registration [7], [8], [9], [10], [11], [12], [13], [14]. For generating CT, these methods can not only relieve the burden of experts but also decrease the exposure time of patients to X-ray radiation, especially for those people who need repeated scans [15], [16]. However, there are still some problems to be resolved.
Firstly, the grayscale differences of images from different modalities are too large, which leads to the poor performance of traditional CNN models when dealing with multi-modal image registration. Secondly, many mainstream methods cannot effectively deal with the deformation of abdominal organs affected by the human body's digestive state and breathing motion. Finally, the performance of DL-based models is highly correlated with the quantity and quality of training dataset. However, it’s a great challenge to collect a large number of paired PET/CT images with high-quality. Therefore, these existing problems motivate us to conduct further research in this area.
In this paper, a multi-resolution registration convolutional neural network (MRR-CNN) is proposed to alleviate these problems. MRR-CNN combined the information of PET and CT in the encoder to learn the features of multi-model images. We also used a mixed loss function including PET and CT to further improve the adaptation performance of the model. Moreover, the proposed multi-resolution structure enables the proposed model to learn large, medium and small deformation of images of different resolutions to predict deformation vector field (DVF) more precisely than the traditional CNN architecture networks. To address the lack of datasets, we proposed a DVF generation strategy for generating new training data.
The overall contributions of this paper are listed as follows:
- •
We proposed a novel registration network to generate DL-based delayed CT (DL-T2CT) images by taking full advantage of existing delayed PET (T2PET), regular PET (T1PET) and regular CT (T1CT) images.
- •
In order to accurately generate DL-T2CT images in delayed scans, the framework of proposed model is designed as MRR-CNN to obtain DVF with multi-scale deformations.
- •
Group normalization (GN) is introduced instead of batch normalization (BN) to reduce the large bias when given a small batch size [17]. GN divides the channels into groups and normalizes within each group to keep its accuracy be stable even for small batch sizes.
- •
A module for simulating delayed scans generation is introduced for training data augmentation.
The rest parts of this paper is organized as follows: Section 2 discusses the related work. Section 3 presents the details of how we built the MRR-CNN. In Section 4, our proposed method is extensively evaluated and provided a performance comparison with other DL-based method. A final conclusion and the potential of the proposed method are presented in Section 5.
Section snippets
Related works
In this section, previous DL-based CT image generation approaches are briefly reviewed. The trend in DL used in the application of medical image registration has been increasing in recent years [7]. The priority of DL-based methods is to realize a significantly accurate and fast prediction of the DVF. VoxelMorph is an unsupervised deformable image registration (DIR) network for MR brain image registration [8]. Then, optional label information is added into VoxelMorph, which make it to be a
Network architecture
As presented in Fig. 1, the proposed MRR-CNN includes three stages to perform registration tasks and generate VF, respectively. T1PET, T2PET and T1CT images are concatenated in the channel scale and then put them into the network,. In the first stage, the concatenated input images are down-sampled by the factor of 4 and then fed into CNN 1 to generate VF 1. We aim to compute large deformations from the low-resolution images in this stage. In the second stage, the concatenated input images are
Data and pre-processing
In this study, clinical scan datasets were acquired using the MinFound ScintCare PET/CT 720L system (Hangzhou, China). The scan of each bed position is continuously acquired within 45 to 60 min after injection of fluoro deoxidized glucose (FDG). PET images were reconstructed by the ordered subset expectation maximization (OSEM) algorithm (2 iterations, 38 subsets). Then, a non-local means (NLM) filter is used for post-smoothing of the reconstruction results [23], [24]. After reconstruction, the
Conclusion
In this paper, a novel registration method based on MRR-CNN was proposed to generate DL-T2CT images using T1CT, T1PET and T2PET images. This multi-resolution strategy enables the proposed method to predict large, medium and small deformations, respectively. Therefore, MRR-CNN can predict DVF more precisely than the traditional CNN architecture networks. The experimental results indicate that the proposed model can achieve more precise prediction results than VoxelMorph.
Through the comprehensive
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This work was partially supported by the National Natural Science Foundation of China (Grant No. 81671038) and the Fundamental Research Funds for the Provincial Universities of Zhejiang (Grant No. GK209907299001-005).
References (26)
- et al.
A diffeomorphic unsupervised method for deformable soft tissue image registration
Comput. Biol. Med.
(2020) - et al.
A novel data augmentation based on Gabor filter and convolutional deep learning for improving the classification of COVID-19 chest X-Ray images
Biomed. Signal. Proces.
(2022) FDG PET/CT: EANM procedure guidelines for tumour imaging: version 2.0
Eur. J. Nucl. Med. Mol. Imaging.
(2015)- et al.
Comparing respiratory gated with delayed scans in the detection of colorectal carcinoma hepatic and pulmonary metastases with 18F-FDG PET-CT
Clin. Nucl. Med.
(2014) Deep learning-based attenuation correction in the absence of structural information for whole-body positron emission tomography imaging
Phys. Med. Biol.
(2020)- et al.
Applications of artificial intelligence in nuclear medicine image generation
Quant. Imaging. Med. Surg.
(2021) - et al.
Multi-Atlas Image Soft Segmentation via Computation of the Expected Label Value, IEEE. Trans. Med
Imaging.
(2021) - et al.
Improved low-dose positron emission tomography image reconstruction using deep learned prior
Phys. Med. Biol.
(2021) - et al.
Deep learning in medical image registration: a review
Phys. Med. Biol.
(2020) - et al.
VoxelMorph: A learning framework for deformable medical image registration, IEEE. Trans. Med
Imaging.
(2019)