Session L30: Novel Acquisition Geometries, Radiation Sources, Hardware, and Algorithms for Medical Imaging

The potential of stationary digital tomosynthesis

Digital Tomosynthesis (DT) as a limited angle 3D imaging technique has already demonstrated its clinical superiority compared to planar 2D X-ray. Current systems however, are expensive, lack mobility and are often susceptible to motion breath artifacts due to long acquisition times. Innovative solutions to miniaturize and basically reinvent the X-ray source such as cold-cathode field emitters, allow for the creation of compact array of individual addressable emitters in the form of a flat panel source (FPS). The advantages of a device using an FPS compared to conventional X-rays and DT systems include the ability to acquire images from different angles without any physical movement, the option to create small and compact devices that can be used for bedside imaging and faster acquisition times that will likely reduce the number and severity of motion artefacts. These advantages will likely result in several clinical benefits that will also be discussed. Potential concerns will also be discussed such as a reduced stand-off-distance of the source to the detector, large incidence angles of x-rays and low signal-to-noise ratio of the individual frames which can lead to various effects such as magnification, scattering, and Poisson noise.   

Cellulose-Based Photonic Nanoparticles for Biomedical Imaging

Here we present a fluorescent targeting nanoparticle contrast agent that can serve as an effective companion solution for biomedical imaging and diagnostics of cancer. Particles can be used to label intra- and extracellular biomarkers and provide key information for clinical decisions based on high resolution, real-time optical imaging while mitigating off-target immunogenic effects. We utilize cellulose acetate to develop a class of benign, natural, and chemically inert fluorescent nanoparticles possessing well-defined safety profiles. Particles are generated from supramolecular assemblies of cellulose acetate and guest polymers, producing composite materials with good biocompatibility, tunable morphology, physical encapsulation ability, and excellent luminescence. We demonstrate effective in vivo targeting of sub-mm tumors in zebrafish cervical cancer xenografts as well as topical targeting of colon cancer tumors in mice. We expect a topically administered contrast agent can provide great clinical value especially in the realm of colorectal endoscopies.   

Deformable motion correction for interventional cone-beam CT

Cone-beam CT (CBCT) is a valuable tool for guiding interventional procedures, including embolization and ablation of soft-tissue targets. However, long scan times (5-30 s) make CBCT susceptible to artifacts arising from involuntary motion. This work reports a method to estimate deformable motion from scan data without additional patient monitoring. A motion vector field (MVF) is computed that minimizes a gradient entropy objective function for image sharpness. MVFs describing deformable motion were estimated as a spatial interpolation of M rigid motion trajectories, each with temporal motion modeled by an N-point spline. Abrupt changes were penalized via spatial-temporal regularization, and a modified 3D filtered backprojection approach was used for motion-corrected image reconstruction. The method was evaluated in digital simulation, cadaver, and retrospective clinical studies. Sharpness of edges at soft-tissue boundaries improved by 75% (4.4±1.1 to 1.1±0.1 mm), 77% (2.2±0.1 mm to 0.5±0.0 mm), and 33% (0.9±0.1 to 0.6±0.1 mm), respectively. These initial studies demonstrate feasibility of correcting deformable organ motion, which will increase the utility and precision of CBCT guidance.   

Quantitative vs qualitative evaluation of automatic segmentation

Automatic segmentation of anatomic regions in medical images has the potential to improve treatment efficiency for image-guided interventions. While many algorithms for automatic segmentation have been developed, evaluation of their clinical usability is largely limited to quantitative metrics such as measures of region overlap (Dice coefficient) or surface distance (Hausdorff distance). Quantitative metrics only tell part of the story; an auto-segmented contour of a small organ, such as an optic nerve, may have a low Dice coefficient but still be clinically acceptable, while a large organ, such as a prostate, may have a high Dice coefficient but be clinically unacceptable. The goal of this work is to explore the use of a qualitative evaluation system for rating the clinical acceptability of auto-segmented contours, and establish the relation between quantitative and qualitative metrics. The qualitative system is designed with 5 levels ranging from “clinically acceptable” to “completely unacceptable” and evaluated by multiple expert observers for pelvic and abdominal structures. If correlation between quantitative and qualitative metrics are found, it would establish scientific basis for the use of quantitative metrics in the evaluation of medical image segmentation.   

A Statistical Model Relating Image Quality to Image Registration Accuracy in Image-Guided Surgery

Image-guided procedures often rely on the ability to accurately register (i.e., align the coordinate systems of) a preoperative image and an intraoperative image. While the accuracy of this registration step is generally thought to increase with improved image quality (in x-ray CT, for example, achieved at the cost of higher dose), there is little quantitative understanding of how registration accuracy relates to image quality. We present a statistical model that relates factors of spatial resolution, noise, and dose to image registration accuracy (viz. root-mean-squared error in the transform parameters). We further show how this framework may be extended to model how rigid registration of bone structures is affected by deformation of surrounding soft-tissue structures. The model is tested in comparison to experiments performed over a range of dose and deformation magnitude showing accurate agreement in general trends and prediction of optimal registration similarity metric. A statistical foundation for understanding the effect of image quality and soft-tissue deformation is an important step in physics-based modeling of imaging systems and guiding the development of new systems for image-guided procedures.   

Neural network-based delineation of clinical target volumes for glioma patients

Outlining the clinical target volume (CTV) in radiotherapy can be time-consuming and error-prone. We propose a convolutional neural network (CNN)-assisted delineation of the CTV for glioma, aiming to reduce inter- and intra-observer variability and decrease treatment planning time. Microscopic disease spread in the brain is restricted by anatomical barriers that are impenetrable by tumor cells. These brain barrier structures were automatically segmented using a 3D CNN trained on 25 datasets of registered planning CT and diagnostic MR images. Satisfactory results were obtained for segmentation of skull, brainstem, corpus callosum, cerebellum, falx cerebri, brain sinuses, tentorium, and venticles. Segmentation quality was assessed by comparing CNN-derived and manually drawn structures using an independent dataset. The Dice score ranged from 73% to 96% and did not improve after more patients were added to the training dataset. After segmentation, the CTV was generated by expanding the gross tumor volume (GTV) by a fixed radius, excluding voxels contained in other segmented structures. We will compare CNN-derived CTV quality with manually delineated CTVs for a large set of 100 patients.
  

Deep Learning Vessel Segmentation for Microsurgical Free Tissue Transfer

Introduction: Free autologous tissue techniques like DIEP (deep inferior epigastric perforator) are regarded as state-of-the-art for patients undergoing breast reconstruction after oncological mastectomy. However, surgeons have to rely on their experience in the identification of the vascular perforators suitable for flap harvest as there exists no standardized approach to this problem.
Objective: To develop a deep-learning method to autonomously segment vessels in the abdominal wall of patients undergoing autologous breast reconstruction to guide pre-surgical planning.
Methods: Manual segmentation of vessel perforators was performed on abdominal CTAs of 24 patients (20 training, 4 validation) undergoing autologous breast reconstruction at the Brigham and Women’s Hospital by an expert rater. A 3D U-net model implemented in DeepNeuro with Keras backend (Beers et al. 2018) was trained to automatically segment vessels in CTAs.
Results: Sensitivity and specificity of the trained model was 0.70/0.70 (training/validation) and 0.99/0.98 (training/validation), respectively.
Conclusion: The described model can reliably perform automatic vessel segmentation in CTA. We will further evaluate this result for guidance of pre-surgical decision making.   

A Deep Learning Approach to Early Cancer Detection using Near-Infrared Laser Scattering Profiles

In the early stages of most cancers, before lesions are visible on a CT or MRI, changes begin to occur at the cellular level as nuclei elongate and mitochondria cluster unevenly. As these organelles are responsible for much (>40%) of the optical scattering which occurs in a cell, changes in cell morphology and structure can largely affect the resulting optical signature. Variations in the physical properties of different cancer types leads to a distinct scattering profile unique to each disease. In this study, optical scattering patterns were investigated from five different cancer cell lines, which were irradiated in vitro with a NIR (854 nm) diode laser. The resulting patterns were collected with a CMOS beam profiler and used to train a convolutional neural network. Differences in these profiles were subtle yet significant enough to allow successful classification by the neural network. After being trained with a set of augmented images from each cancer type, the network was able to distinguish cell lines with an accuracy of up to 98.5%. The accurate classification of these patterns at low concentrations could contribute to the early detection of cancerous cells in otherwise healthy tissue. Current methods will also be discussed such as semantics and instance segmentation.   

Quantitative Cone-Beam CT with High-fidelity Modeling of Imaging Physics

Bone mineral density (BMD) and bone microstructure are key biomarkers of orthopedic health. Quantitative assessment of these parameters requires high accuracy of reconstructed attenuation values and high spatial resolution. We employ advanced models of x-ray propagation to optimize performance of specialized orthopedic Cone-Beam CT systems in quantitative bone imaging. To achieve accurate measurements of BMD, a model-based reconstruction (MBR) framework utilizing polyenergetic spectral models is used in concert with fast Monte Carlo scatter correction. To advance spatial resolution to a level consistent with trabecular detail (~100 μm), we adopt a customized low-noise CMOS x-ray detector with 400 μm-thick scintillator, optimized through cascaded systems modeling of task-based imaging performance. Studies of BMD accuracy indicate that MBR is able to estimate CaCO₃ concentration with <20 mg/mL error, irrespective of object size and position. The use of optimized CMOS sensor yielded improved correlation with gold-standard micro-CT measurements of bone microstructure compared to current-generation flat-panel detector CBCT, e.g. correlation for trabecular thickness increased from 0.84 with a flat-panel CBCT to 0.96 with CMOS.   

Quantitative evaluation of inflammatory response dynamics in the lung following proton and photon irradiation

We analyzed lung density changes in lung cancer patients receiving stereotactic body radiation therapy with protons (SBPT) or photons (SBRT). Follow-up computer tomography (CT) scans were registered to pre-treatment scans using B-spline based deformable image registration. Dose response curves (DRC) were used to correlate the radiographic change in Hounsfield Units (HU) to the radiation dose and fitted using linear regression to provide a quantitative measure of normal lung response. CTs were also evaluated by a thoracic radiologist.
46 patients (23 SBPT/SBRT) were matched based on patient characteristics, quantitative assessment correlated highly with expert evaluation (p<10^-13,rho=0.6). The maximum response averages 3.8±3.0 (SBPT) versus 3.2±2.4 HU/Gy (SBRT). The lung response after SBRT significantly increases over the observation period (p=0.003), but not after SBPT, indicating an accelerated acute response after SBPT.
We demonstrate that quantitative normal lung response evaluation correlates well with established grading systems and differences in response dynamics after SBPT compared to SBRT, that warrant further investigation into their mechanisms of inflammation.   

Measurement of In vitro Cancer Tumor Hypoxia

Tumors are characterized as swamps: abnormal and disordered tissue masses with highly stressful conditions of hypoxia, low pH, low nutrient conditions due to a combination of rapid cell growth, lack of vasculature and altered metabolism. While for normal cells that combination would be lethal, for cancer cells it provides a genotoxic environment they are adapted to. We show here using a phosphorescence lifetime imaging (PLIM) technology based oxygen sensor to monitor the local O$_{2}$ level in a extended two dimensional array of cancer cells with strong and mixed gradients to nutrients and O$_{2}$ using a novel pure diffusional three dimensional microfabricated technology the emergence of highly hypotoxic dormant cell metapopulations.
  

Novel X-ray Sources for Medical Imaging: Making Old Physics Do New Tricks

The diagnosis of medical conditions often relies on medical imaging. Imaging applies basic physics and the techniques deployed have often found their origin in experimental methods. While some of these imaging modalities are new and have rapidly evolved, x-ray imaging has remained relatively stagnant since the introduction of Computed Tomography (CT) and more recently digital x-ray detectors. The availability of low-noise detectors, high speed desk-top computers and new algorithms offers incremental improvements on established x-ray imaging, but also challenges the practitioners in optimally deploying these resources as a primary concern is patient lifetime dose. New x-ray sources are often needed for advanced imaging approaches such as digital tomosynthesis, phase contrast imaging or improved interventional radiology. These novel sources often rely on field enhanced emission and carry new challenges into clinically deployed devices including ultra high vacuum, high voltage switching and the need for active feedback controls. I will describe some of these challenges, the regimes of operation under consideration by various groups, and the practical implications of the physics parameters to radiology.