Advanced Image Formation Algorithms in Tomographic X-ray Imaging – Invited Special Session

C2L-D: Advanced Image Formation Algorithms in Tomographic X-ray Imaging - Invited Special Session

Session Type: Lecture
Session Code: C2L-D
Location: Room 4
Date & Time: Friday March 24, 2023 (10:00-11:00)
Chair: Alejandro Sisniega
Track: 12
Paper IDPaper TitleAuthorsAbstract
3152Statistical Deep Learning for CT Image Formation with Tunable Control of Signal and NoiseMatthew Tivnan{2}, Jacopo Teneggi{2}, Yijie Yuan{2}, Tzu-Cheng Lee{1}, Ruoqiao Zhang{1}, Kirsten Boedeker{1}, Liang Cai{1}, Jeremias Sulam{2}, J. Webster Stayman{2}Medical image reconstruction and restoration are some of the most high-stakes applications of new deep learning models and methods. On the positive side, improvements to medical image quality can help radiologists and other clinicians to detect, diagnose, and take action to address critical medical issues. Conversely, the errors and artifacts introduced by these models can have a real-world negative impact on patient health outcomes. Therefore, there is a critical need for a greater level of understanding and control over the errors in the outputs of deep learning image formation models. In this work, we focus on the specific clinical task of low radiation dose CT image restoration and we propose a new deep learning model architecture for tunable neural networks with auxiliary inputs to control the spatial resolution, noise texture, and noise magnitude. During training, we use multiple realizations of the noisy blurry input images leading to multiple realizations of the deep learning output images. We propose a KL divergence loss function between this output distribution and a target distribution parameterized by the same spatial resolution, noise texture, and noise magnitude coefficients provided as auxiliary inputs. Our results demonstrate the novel capability for tunable control of image quality with auxiliary inputs. We show that the output spatial resolution and noise properties from our method are highly predictable, especially for modest targeted noise reduction and sharpening. Finally, we demonstrate that our method can be used to optimize image quality for diagnosing the malignancy of lung nodules based on change in shape.
3176Deep Autofocus for Deformable Motion Compensation in Interventional CBCT via Context-Aware Metrics and Adaptive Temporal SamplingHeyuan Huang{1}, Alex Lu{1}, Yicheng Hu{1}, Wojciech Zbijewski{1}, Mathias Unberath{1}, Clifford Weiss{1}, Jeffrey Siewerdsen{2}, Alejandro Sisniega{1}Cone-beam CT (CBCT) is ubiquitous in the interventional suite for three-dimensional visualization of fine vascularity in abdominal interventions, but long acquisition time makes it susceptible to patient motion. Limited soft-tissue contrast, coupled to the deformable and non-periodic nature of abdominal motion, challenge image-based motion compensation methods with limitations to simultaneously quantify motion artifacts and anatomical integrity and to accommodate complex spatial-temporal motion patterns within the optimization process. We propose a metric for simultaneous quantification of motion severity and anatomical realism via a reference-free learned similarity metric trained to replicate Visual Information Fidelity (VIF) in presence of CBCT motion. Quantification of local motion is achieved with a context aware deep convolutional network combining a high-resolution branch that provides quantification of local motion severity in the region of interest with a low-resolution branch that extract features associated to the global spatial context. Deformable motion compensation is achieved via integration of the context- aware autofocus metric into a four-dimensional spatial transformer network acting on the evolving CBCT volume. The motion trajectory estimation is posed as a multi-stage problem using Hermite splines with adaptive amplitude and temporal positioning, allowing adaption to non-periodic irregular motion within a fully differentiable framework. Deformable motion compensation was demonstrated in simulated abdominal CBCT data, including realistic models of the imaging chain, and in clinical images obtained for guidance of transarterial chemoembolization in the liver. The deep autofocus approach yielded mitigation of artifacts and restoration of the underlying anatomy, outperforming conventional autofocus based of image sharpness metrics and fixed sampling strategies.
3184Quantitative Dual-Energy Cone-Beam CT Using Optimization-Based Image Formation AlgorithmsStephen Liu, Alejandro Sisniega, J. Webster Stayman, Wojciech ZbijewskiDual-Energy (DE) x-ray imaging techniques enable improved material discrimination compared to conventional single-energy approaches. However, DE material decomposition is typically susceptible to noise amplification. Further, its quantitative accuracy might be challenged by residual biases and non-idealities in the x-ray projection data. We present model-based nonlinear approaches to address these challenges. The focus of this work is on a novel point-of-care Cone-Beam CT (CBCT) for orthopedic applications. The scanner uses a unique multi-source x-ray tube to achieve fast (<10s data acquisition) low-dose single-scan DE imaging. X-ray scatter - the primary source of projection bias in CBCT - is corrected using fast Monte Carlo (MC) simulations. The MC scatter model and a polyenergetic model of primary radiation are incorporated in a nonlinear optimization algorithm to reconstruct the measured DE data into 3D concentration maps of three compounds of interest (e.g., calcium, water, adipose in bone imaging). The key to this approach is the integration of volume conservation principle (VCP), which requires that material fractions (nonnegative) in each reconstructed voxel add to unity, as a pair of local constraints in the optimization problem. In this manner, we are able to reconstruct more than two materials from the two DE measurements. Other notable features of the algorithm include reduction of metal-induced image artifacts, achieved by either incorporating a shape model of the metal implant (if available) as a prior, or by appropriately adjusting the base material sets in the regions under the VCP constraint. The proposed approach was validated on physical data acquired from the extremity CBCT scanner (available at Johns Hopkins Hospital), and its performance was compared to various conventional DE algorithms in the measurements of bone mineral density and detection of bone marrow edema.
3230Uncertainty Quantification in CT DenoisingJacopo Teneggi, Jeremias SulamModern algorithmic tools have allowed for the restoration of degraded images to a remarkable degree, enabling excellent denoising results in computed tomography. Yet, all of these accurate reconstructions represent point estimates, and algorithms do not -in general- provide any information on the uncertainty surrounding these estimates. In this talk, we will show how to tie modern diffusion-based models for image restoration to conformal prediction, resulting in precise calibrated uncertainty intervals together with representative samples from the corresponding posterior distribution given the measurements. We will illustrate these methods for CT denoising, comment on current limitations and remaining open questions.