BCD-Net for Low-Dose CT Reconstruction: Acceleration, Convergence, and Generalization

Abstract

Obtaining accurate and reliable images from low-dose computed tomography (CT) is challenging. Regression convolutional neural network (CNN) models that are learned from training data are increasingly gaining attention in low-dose CT reconstruction. This paper modifies the architecture of an iterative regression CNN, BCD-Net, for fast, stable, and accurate low-dose CT reconstruction, and presents the convergence property of the modified BCD-Net. Numerical results with phantom data show that applying faster numerical solvers to modelbased image reconstruction (MBIR) modules of BCD-Net leads to faster and more accurate BCD-Net; BCD-Net significantly improves the reconstruction accuracy, compared to the state-of-the-art MBIR method using learned transforms; BCD-Net achieves better image quality, compared to a state-of-the-art iterative NN architecture, ADMM-Net. Numerical results with clinical data show that BCD-Net generalizes significantly better than a state-of-the-art deep (non-iterative) regression NN, FBP-ConvNet, that lacks MBIR modules. The authors indicated by asterisks ( * ) equally contributed to this work. Corresponding author: Yong Long (email: yong.long@sjtu.edu.cn). This paper has supplementary document. The prefix "S" indicates the numbers in figure and section in the supplementary document. arXiv:1908.01287v1 [eess.IV] 4 Aug 2019 2 I.Y. Chun and X. Zheng et al.solving large-scale inverse problems in imaging, the first scheme is limited in training CNNs from large-scale images; the second scheme does not effectively remove complicated noise features; and the third scheme has limited benefits when applied to convolutional layers.An alternative way to regulate overfitting of regression CNNs in inverse imaging problems is combining them with model-based image reconstruction (MBIR) that considers imaging physics or image formation models, and noise statistics in measurements. is an iterative regression CNN that generalizes a block coordinate descent (BCD) MBIR method using learned convolutional regularizers [5]. Each layer (or iteration) of BCD-Net consists of image denoising and MBIR modules. In particular, the denoising modules use layer-wise regression CNNs to effectively remove layer-dependent noise features. Many existing works can be viewed as a special case of BCD-Net. For example, RED [11] and MoDL [1] are special cases of BCD-Net, because they use identical image denoising modules across layers or only consider quadratic data-fidelity terms (e.g., the first term in (P1)) in their MBIR modules.This paper modifies BCD-Net that uses convolutional autoencoders in its denoising modules [4], and applies the modified BCD-Net to low-dose CT reconstruction. First, for fast CT reconstruction, we apply the Accelerated Proximal Gradient method using a Majorizer (APG-M), e.g., FISTA [2], to MBIR modules using the statistical CT data-fidelity term. Second, this paper provides the sequence convergence guarantee of BCD-Net when applied to low-dose CT reconstruction. Third, it investigates the generalization ca...

Improved Low-Count Quantitative PET Reconstruction With an Iterative Neural Network

“…BCD-Net is inspired by this type of "learned" regularizer. Ultimately, we hope that the learned regularizer can better separate true signal from noisy components compared to hand-crafted filters [29].…”

“…Iterative NNs [8]- [11], [21]- [24] are a broad family of methods that originate from an unrolling algorithm for solving an optimization problem and BCD-Net [25] is a specific example of an iterative NN. BCD-Net is constructed by unfolding a block coordinate descent (BCD) MBIR algorithm using "learned" convolutional analysis operators [26]- [28], leading to significantly improved image recovery accuracy in extreme imaging applications, e.g., low-dose CT [29], dual-energy CT [30], highly undersampled MRI [25], denoising low-SNR images [25], etc. A preliminary version of this paper was presented at the 2018 Nuclear Science Symposium and Medical Imaging Conference [31].…”

A Review of Deep Learning Methods for Compressed Sensing Image Reconstruction and Its Medical Applications

“…That is to say, both P image s and P dc s are represented by neural networks. The authors of [86] used a denoising auto-encoder with soft-thresholding function as P image and solved the P dc by FISTA. In each stage, a cleaner image z is obtained by the denoising model and FISTA is used to keep data fidelity of z.…”

“…FBPConvNet is chosen as the neural network structure and PWLS-EP or PWLS-ULTRA is the choice for the iterative algorithm. Similar to [86], the training scheme is a stage-wise process.…”

Advances in micro-CT imaging of small animals

“…Variational models borrow the structure of these proven techniques, guiding stochastic training with domain-specific priors, while maintaining the myriad benefits of data-driven generalization. Successful application of variational models has been demonstrated in both MRI [107] and CT [96,97,99,100,108] denoising and reconstruction tasks. As we will illustrate in the following sub-sections on other DL applications, the incorporation of domain-specific knowledge into DL projects is a clear trend and will be critical to the adoption of DL technology in routine practice.…”