Yixing Huang

Some Investigations on Robustness of Deep Learning in Limited Angle Tomography

In computed tomography, image reconstruction from an insufficient angular range of projection data is called limited angle tomography. Due to missing data, reconstructed images suffer from artifacts, which cause boundary distortion, edge blurring, and intensity biases. Recently, deep learning methods have been applied very successfully to this problem in simulation studies. However, the robustness of neural networks for clinical applications is still a concern. It is reported that most neural networks are vulnerable to adversarial examples. In this paper, we aim to investigate whether some perturbations or noise will mislead a neural network to fail to detect an existing lesion. Our experiments demonstrate that the trained neural network, specifically the U-Net, is sensitive to Poisson noise. While the observed images appear artifact-free, anatomical structures may be located at wrong positions, e.g. the skin shifted by up to 1 cm. This kind of behavior can be reduced by retraining on data with simulated Poisson noise. However, we demonstrate that the retrained U-Net model is still susceptible to adversarial examples. We conclude the paper with suggestions towards robust deep-learning-based reconstruction.

Make the Most of Time Temporal Extension of the iTV Algorithm for 4D Cardiac C-Arm CT

Gated 4D cardiac imaging with C-arm CT scanners suffers from insufficient image quality due to strong angular undersampling. To deal with this problem, we suggest an iterative reconstruction method with spatial and temporal total variation regularization based on an established framework which controls the relative contributions of raw data error minimization and regularization. This new method is tested on a simulated heart phantom and on two clinical data sets. We show that the additional use of temporal regularization is advantageous compared to spatial regularization exclusively, with the relative root mean square error lowered from 11.75% to 8.24% in the phantom study.

Traditional Machine Learning Techniques for Streak Artifact Reduction in Limited Angle Tomography

In this work, the application of traditional machine learning techniques, in the form of regression models based on conventional, “hand-crafted” features, to streak reduction in limited angle tomography is investigated. Specifically, linear regression (LR), multi-layer perceptron (MLP), and reduced-error pruning tree (REPTree) are investigated. When choosing the mean-variation-median (MVM), Laplacian, and Hessian features, REPTree learns streak artifacts best and reaches the smallest root-mean-square error (RMSE) of 29HU for the Shepp-Logan phantom. Further experiments demonstrate that the MVM and Hessian features complement each other, whereas the Laplacian feature is redundant in the presence of MVM. Preliminary experiments on clinical data suggests that further investigation of clinical applications using REPTree may be worthwhile.

Overexposure Correction by Mixed One-Bit Compressive Sensing for C-Arm CT

This paper proposes a novel method to deal with overexposure for C-arm CT reconstruction. The proposed method is based on recent progress of one bit compressive sensing (1bit-CS), which is to recover sparse signals from sign measurements. Overexposure could be regarded as a kind of sign information, thus the application of 1bit-CS to overexposure correction in CT reconstruction is expected. This method is evaluated on a phantom and its promising performance implies potential application on clinical data.

Comparison of SART and ETV reconstruction for increased C-arm CT volume coverage by proper detector rotation in liver imaging

In this work, we present a method to increase the lateral field-of-view of a C-arm CT system by rotating the detector such that the diagonal of the detector lies on the u-axis of the detectors coordinate system. We investigated three different 3-D scan trajectories for liver imaging of an obese patient (waist circumference 130 cm) — a Short Scan, a Large Volume Scan and a Helical Large Volume Scan. We reconstructed a data set of the Visible Human Project with the SART and the eTV algorithm. Tests revealed that the coverage was increased with the presented method by 25.3 % for the Short Scan and 28.5 % for the Large Volume Scan. Performing helical scans compensated the axial data loss. The two implemented iterative approaches both provide acceptable results, with the eTV algorithm reducing the RMSE compared to SART by about 29 %. Given a liver imaging task, the rotated detector is able to image the entire liver section of the abdomen with a single Large Volume Scan.

Image Quality Analysis of Limited Angle Tomography Using the Shift-Variant Data Loss Model

This paper investigates the application of the shift-variant data loss (SVDL) model in image quality assessment for a state-ofthe-art reconstruction technique, the weighted total variation (wTV), in limited angle tomography. The SVDL model is used to analyze the acquired frequency information in 2-D fan-beam limited angle tomography. The wTV algorithm is applied to reconstruct some specific mathematical phantoms. The experiments show that the reconstructed image quality depends on the relation of the source trajectory and geometric structure of the imaged object, position, shape, size and orientation in particular.