Research

While the AIAI lab is involved in many research projects, there are several overarching themes that have tied much of the research together. Some of these major themes are listed below:

.

Machine/Deep Learning in Imaging

Combining Physical Measurement Models and Deep Learning: Diffusion Posterior Sampling Neural Networks Designed with Multiple Noise Realizations and a Noise vs. Bias Loss Function A tunable neural network for CT denoising Generative adversarial networks and radiomics supervision for lung lesion synthesis Manifold reconstruction of difference: MBIR with a data-driven prior

Machine learning is finding widespread application including imaging applications. In particular, we are interested in learned approaches that can be applied as part of the image formation process. Moreover, we seek machine learning solutions that leverage physics and domain knowledge about imaging performance and clinical tasks as part of their design and training. Similarly, we are interested in the rigorous assessment of machine learning approaches in terms of task-based performance, robustness, and generalization. Applications are broad but include: Parameterized neural networks that can provide task-optimized performance or a range of image properties with a user-tunable “knob” to permit, e.g., the level of noise reduction; Networks that use radiomics supervision to create realistic anatomy for virtual clinical trials and algorithm assessments; Networks that combine rigorous physics models with deep learned priors of human anatomy; Perturbation analysis of machine learning denoising approaches. Selected publications include:

.

Spectral Computed Tomography

Spectral CT using spatial-spectral filters Spectral Diffusion Posterior Sampling for Material Decomposition Grating/grid-based spectral CT Enhancement of spectral CT performance using combination approaches Dual-layer flat-panel, high-resolution, spectral CT with layer-dependent blur modeling

Standard computed tomography does not leverage energy-dependent attenuation of materials; however, there is growing use of this information for spectral CT which permits material identification and decomposition with quantitative density estimates. While several dual-energy approaches have found clinical application including kV-switching, dual-layer detectors, and dual-source systems, there is room for additional strategies. We have developed several novel strategies including spatial-spectral filtration, grating-based filtration, as well as combination approaches for improved performance. Enabling these new data acquisition methods and their combinations are model-based material decomposition methods (e.g. so-called “one-step” approaches) that implicitly handle highly varied measurement models, sparsity, etc. Optimization and control of such algorithms is potentially complex and has necessitated studies on how to control not only the standard noise-resolution trade-offs but also the cross-response and correlations between channels for optimal regularization and system design. Selected publications include:

.

Virtual Clinical Trials and Phantom Studies

Anthropomorphic Phantom 3D Printing based on Patient Scans Pelvis Phantom with 3D-Printed Prostate Volume and Bilateral Hip Implants Procedurally generated texture for investigation of CT imaging performance

Testing new medical imaging technologies can be very costly when it involves patient studies. Thus, the ability to conduct simulation studies and physical phantom experiments is often a critical initial step in the introduction of new technologies. This is particularly true for new system design and algorithm development. Morever, there is growing agreement that systematic evaluation of imaging systems needs to involve performance metrics that quantify the clinical task that is being conducted. Thus, for realistic studies, virtual clinical trials and phantom experiments should represent actual clinical tasks and realistic anatomical structures. Towards this end, we have developed a number of simulation tools and processes for phantom construction of realistic anthropomorphic targets. Selected publications include:

Task-Driven Imaging

Task-Driven Tube Current Modulation in Low-Dose Abdominal CT Task-Driven Trajectories Reconstructions for Task-Driven Trajectories for Embolized Aneurysms Orbits for viewing around metal implants Customized orbits for viewing around metal implants

Traditional image quality metrics like noise variance and RMS error often fail to capture details of imaging performance. More sophisticated performance metric integrate a mathematical formulation of the diagnostic task to quantify the expected task-based performance. Such metrics have been used extensively for system design and evaluation; however, one can go a step further and use such metrics to prospectively predict imaging performance under different acquisition and reconstruction settings. Thus, it is possible to perform task-driven imaging that is optimized for the particular diagnostic test and patient anatomy. We have developed a optimization framework that can predict imaging performance using detectability index to drive customized acquisitions and image formation for both traditional filtered-backprojection and model-based reconstruction methods. Such optimization can lead to novel acquisition strategies. For example, for computed tomography with tube current modulation (TCM) we find scenarios where current TCM strategies can degrade task performance when model-based reconstruction is applied. In cone-beam CT applications we have demonstrated that if one breaks free from traditional circular or helical orbits, dramatic image quality improvement can be made when more general non-circular source-detector trajectories are chosen. Selected publications include:

.

Patient-Specific Data Acquisition

Customized spectral shaping using fine pitch lamellae Noncircular orbits on a floor mounted robotic C-arm Fluence-field modulation in CT using multiple aperture devices Customized orbits for interventional imaging Patient-specific source-detector trajectories on a CBCT test-bench

With the capability to optimize patient imaging through the task-driven imaging framework discussed above, we have sought flexibility in imaging system designs. Thus, we have built imaging platforms with expanded capabilities to customize the data acquisition. This includes a modified cone-beam CT test-bench with the capability to perform frame-to-frame fluence modulation up to a dynamic range of about 60 (in contrast, most modern clinical CT scanners are capable of dynamic tube current modulation of around a factor of three). Similarly, the AIAI lab has developed fluence-field modulation strategies (analogous to dynamic bowtie filters) that permit varying spatial modulations of the x-ray beam during an acquisition. Alternate strategies using similar technology have allow for customization of the x-ray spectrum. Additionally, we have developed and investigated test-bench and clinical systems that are capable of generalized non-circular orbits (in contrast to traditional CT universally uses a circular orbit). Such systems have much more flexibility in selecting what information is collected during a patient scan and how much and where x-ray exposure is delivered – permitting highly patient-specific data acquisition where imaging performance is maximized while minimizing radiation exposure. Selected publications include:

.

High-Fidelity System Modeling

Iterative reconstruction with system blur and correlation models (GPL-BC) Modeling post-correction noise variance in CBCT measurements Shift-variant blur due to different apparent focal spot size/shape across the detector

Model-based Iterative Reconstruction (MBIR) has been widely adopted in tomographic applications due to its ability to model imaging system details and measurement noise. The limits of imaging performance using such methods rely on accurate imaging system models – in effect, the higher the fidelity of the system model, the better the imaging performance that can be achieved. Many MBIR methods have been applied to imaging systems without careful consideration of the underlying physical models. This has been particularly true for flat-panel-based cone-beam CT systems where traditional assumptions about the physical model are violated. For example, conventional MBIR methods presume independent measurement and no blur between detector elements; whereas such effects can be prominent in high-resolution flat-panel CT systems. We have developed MBIR models that integrate high-fidelity models to demonstrate improved imaging performance. Related work has consider other variations in the noise model based on various corrections that are applied to CBCT projection data which also change the noise structure. Selected publications include:

.

Prior-Image-Based Reconstruction

dPIRPLE Reconstruction for ultra-low dose lung nodule surveillance dPIRPLE: Alternating image and registration updates Direct Reconstruction of Difference (RoD) for estimation of anatomical change

There are numerous scenarios where patients undergo a series of imaging studies. For example, sequential studies are often obtained over the course of a disease diagnosis, treatment, and follow-up. Specific clinical scenarios include dynamic imaging studies (perfusion imaging, image-guided surgery with intraoperative imaging, image-guided radiotherapy with on-board imaging, etc.) Such sequential studies share a great wealth of shared anatomical information that can potentially be used to reduce data fidelity requirements in subsequent scans (permitting faster, lower exposure, etc. acquisitions). We have developed a number of reconstruction methods that utilize this highly specific prior image information as part of the image formation process. Clinical examples have included very low exposure lung nodule surveillance using sparse acquisitions, and improved onboard CBCT imaging for radiation therapy using CT planning data. Some methods like Reconstruction of Difference (RoD) have focused on imaging the change between studies; whereas the PIRPLE (Prior Image Registration Penalized Likelihood Estimation) approach has concentrated on visualization of the current anatomy. These prior image methods utilize high-fidelity physical models including measurement statistics for optimal performance and have been used jointly with both rigid and deformable registration of the prior image to better integrate prior information. Selected publications include:

.

Known-Component Reconstruction

Reconstruction Comparison - CBCT Experiments KCR - Alternating Registration and Image Updates Reconstruction Comparison - Cadaver Data with Pedicle Screw

Interventional imaging presents a number of opportunities to integrate additional knowledge into the image formation process. This is particularly true in 3D imaging of implants within a patient. In CT, metal implants can cause significant artifacts including noise and streaking that degrade the image. Moreover, the image quality tends to worst nearest the implant – which is typically where image quality requirements are greatest (i.e., diagnostic tasks are typically close to the implant/tissue boundary). However, if the shape and material composition of the implant is known a priori (e.g., from a CAD model), one may integrate this information into image formation – effectively creating a joint reconstruction problem (of the anatomical background) and a registration of the known component(s). We have applied such methodologies for a number of styles of implant and generalized the approach for deformable and other generalized registrations; eliminated the requirement for prior knowledge of beam quality and material composition through joint estimation of spectral effects; and removed the limitation of needing a shape model by deriving shape information directly from the projection data. Selected publications include: