In this work, we present approximate adders and multipliers to reduce data-path complexity of specialized hardware for various image processing systems. These approximate circuits have a lower area, latency and power consumption compared to their accurate counterparts and produce fairly accurate results. We build upon the work on approximate adders and multipliers presented in [23] and [24]. First, we show how choice of algorithm and parallel adder design can be used to implement 2D Discrete Cosine Transform (DCT) algorithm with good performance but low area. Our implementation of the 2D DCT has comparable PSNR performance with respect to the algorithm presented in [23] with ~35-50% reduction in area. Next, we use the approximate 2x2 multiplier presented in [24] to implement parallel approximate multipliers. We demonstrate that if some of the 2x2 multipliers in the design of the parallel multiplier are accurate, the accuracy of the multiplier improves significantly, especially when two large numbers are multiplied. We choose Gaussian FIR Filter and Fast Fourier Transform (FFT) algorithms to illustrate the efficacy of our proposed approximate multiplier. We show that application of the proposed approximate multiplier improves the PSNR performance of 32x32 FFT implementation by 4.7 dB compared to the implementation using the approximate multiplier described in [24]. We also implement a state-of-the-art image enlargement algorithm, namely Segment Adaptive Gradient Angle (SAGA) [29], in hardware. The algorithm is mapped to pipelined hardware blocks and we synthesized the design using 90 nm technology. We show that a 64x64 image can be processed in 496.48 µs when clocked at 100 MHz. The average PSNR performance of our implementation using accurate parallel adders and multipliers is 31.33 dB and that using approximate parallel adders and multipliers is 30.86 dB, when evaluated against the original image. The PSNR performance of both designs is comparable to the performance of the double precision floating point MATLAB implementation of the algorithm.

Coronary computed tomography angiography (CTA) has a high negative predictive value for ruling out coronary artery disease with non-invasive evaluation of the coronary arteries. My work has attempted to provide metrics that could increase the positive predictive value of coronary CTA through the use of dual energy CTA imaging. After developing an algorithm for obtaining calcium scores from a CTA exam, a dual energy CTA exam was performed on patients at dose levels equivalent to levels for single energy CTA with a calcium scoring exam. Calcium Agatston scores obtained from the dual energy CTA exam were within ±11% of scores obtained with conventional calcium scoring exams. In the presence of highly attenuating coronary calcium plaques, the virtual non-calcium images obtained with dual energy CTA were able to successfully measure percent coronary stenosis within 5% of known stenosis values, which is not possible with single energy CTA images due to the presence of the calcium blooming artifact. After fabricating an anthropomorphic beating heart phantom with coronary plaques, characterization of soft plaque vulnerability to rupture or erosion was demonstrated with measurements of the distance from soft plaque to aortic ostium, percent stenosis, and percent lipid volume in soft plaque. A classification model was developed, with training data from the beating heart phantom and plaques, which utilized support vector machines to classify coronary soft plaque pixels as lipid or fibrous. Lipid versus fibrous classification with single energy CTA images exhibited a 17% error while dual energy CTA images in the classification model developed here only exhibited a 4% error. Combining the calcium blooming correction and the percent lipid volume methods developed in this work will provide physicians with metrics for increasing the positive predictive value of coronary CTA as well as expanding the use of coronary CTA to patients with highly attenuating calcium plaques.

Magnetic Resonance Imaging using spiral trajectories has many advantages in speed, efficiency in data-acquistion and robustness to motion and flow related artifacts. The increase in sampling speed, however, requires high performance of the gradient system. Hardware inaccuracies from system delays and eddy currents can cause spatial and temporal distortions in the encoding gradient waveforms. This causes sampling discrepancies between the actual and the ideal k-space trajectory. Reconstruction assuming an ideal trajectory can result in shading and blurring artifacts in spiral images. Current methods to estimate such hardware errors require many modifications to the pulse sequence, phantom measurements or specialized hardware. This work presents a new method to estimate time-varying system delays for spiral-based trajectories. It requires a minor modification of a conventional stack-of-spirals sequence and analyzes data collected on three orthogonal cylinders. The method is fast, robust to off-resonance effects, requires no phantom measurements or specialized hardware and estimate variable system delays for the three gradient channels over the data-sampling period. The initial results are presented for acquired phantom and in-vivo data, which show a substantial reduction in the artifacts and improvement in the image quality.

Locomotion of microorganisms is commonly observed in nature and some aspects of their motion can be replicated by synthetic motors. Synthetic motors rely on a variety of propulsion mechanisms including auto-diffusiophoresis, auto-electrophoresis, and bubble generation. Regardless of the source of the locomotion, the motion of any motor can be characterized by the translational and rotational velocity and effective diffusivity. In a uniform environment the long-time motion of a motor can be fully characterized by the effective diffusivity. In this work it is shown that when motors possess both translational and rotational velocity the motor transitions from a short-time diffusivity to a long-time diffusivity at a time of pi/w. The short-time diffusivities are two to three orders of magnitude larger than the diffusivity of a Brownian sphere of the same size, increase linearly with concentration, and scale as v^2/2w. The measured long-time diffusivities are five times lower than the short-time diffusivities, scale as v^2/{2Dr [1 + (w/Dr )^2]}, and exhibit a maximum as a function of concentration. The variation of a colloid's velocity and effective diffusivity to its local environment (e.g. fuel concentration) suggests that the motors can accumulate in a bounded system, analogous to biological chemokinesis. Chemokinesis of organisms is the non-uniform equilibrium concentration that arises from a bounded random walk of swimming organisms in a chemical concentration gradient. In non-swimming organisms we term this response diffusiokinesis. We show that particles that migrate only by Brownian thermal motion are capable of achieving non-uniform pseudo equilibrium distribution in a diffusivity gradient. The concentration is a result of a bounded random-walk process where at any given time a larger percentage of particles can be found in the regions of low diffusivity than in regions of high diffusivity. Individual particles are not trapped in any given region but at equilibrium the net flux between regions is zero. For Brownian particles the gradient in diffusivity is achieved by creating a viscosity gradient in a microfluidic device. The distribution of the particles is described by the Fokker-Planck equation for variable diffusivity. The strength of the probe concentration gradient is proportional to the strength of the diffusivity gradient and inversely proportional to the mean probe diffusivity in the channel in accordance with the no flux condition at steady state. This suggests that Brownian colloids, natural or synthetic, will concentrate in a bounded system in response to a gradient in diffusivity and that the magnitude of the response is proportional to the magnitude of the gradient in diffusivity divided by the mean diffusivity in the channel.

A cerebral aneurysm is a bulging of a blood vessel in the brain. Aneurysmal rupture affects 25,000 people each year and is associated with a 45% mortality rate. Therefore, it is critically important to treat cerebral aneurysms effectively before they rupture. Endovascular coiling is the most effective treatment for cerebral aneurysms. During coiling process, series of metallic coils are deployed into the aneurysmal sack with the intent of reaching a sufficient packing density (PD). Coils packing can facilitate thrombus formation and help seal off the aneurysm from circulation over time. While coiling is effective, high rates of treatment failure have been associated with basilar tip aneurysms (BTAs). Treatment failure may be related to geometrical features of the aneurysm. The purpose of this study was to investigate the influence of dome size, parent vessel (PV) angle, and PD on post-treatment aneurysmal hemodynamics using both computational fluid dynamics (CFD) and particle image velocimetry (PIV). Flows in four idealized BTA models with a combination of dome sizes and two different PV angles were simulated using CFD and then validated against PIV data. Percent reductions in post-treatment aneurysmal velocity and cross-neck (CN) flow as well as percent coverage of low wall shear stress (WSS) area were analyzed. In all models, aneurysmal velocity and CN flow decreased after coiling, while low WSS area increased. However, with increasing PD, further reductions were observed in aneurysmal velocity and CN flow, but minimal changes were observed in low WSS area. Overall, coil PD had the greatest impact while dome size has greater impact than PV angle on aneurysmal hemodynamics. These findings lead to a conclusion that combinations of treatment goals and geometric factor may play key roles in coil embolization treatment outcomes, and support that different treatment timing may be a critical factor in treatment optimization.

Controlled release formulations for local, in vivo drug delivery are of growing interest to device manufacturers, research scientists, and clinicians; however, most research characterizing controlled release formulations occurs in vitro because the spatial and temporal distribution of drug delivery is difficult to measure in vivo. In this work, in vivo magnetic resonance imaging (MRI) of local drug delivery is performed to visualize and quantify the time resolved distribution of MRI contrast agents. I find it is possible to visualize contrast agent distributions in near real time from local delivery vehicles using MRI. Three dimensional T1 maps are processed to produce in vivo concentration maps of contrast agent for individual animal models. The method for obtaining concentration maps is analyzed to estimate errors introduced at various steps in the process. The method is used to evaluate different controlled release vehicles, vehicle placement, and type of surgical wound in rabbits as a model for antimicrobial delivery to orthopaedic infection sites. I are able to see differences between all these factors; however, all images show that contrast agent remains fairly local to the wound site and do not distribute to tissues far from the implant in therapeutic concentrations. I also produce a mathematical model that investigates important mechanisms in the transport of antimicrobials in a wound environment. It is determined from both the images and the mathematical model that antimicrobial distribution in an orthopaedic wounds is dependent on both diffusive and convective mechanisms. Furthermore, I began development of MRI visible therapeutic agents to examine active drug distributions. I hypothesize that this work can be developed into a non-invasive, patient specific, clinical tool to evaluate the success of interventional procedures using local drug delivery vehicles.

Ultrasound imaging is one of the major medical imaging modalities. It is cheap, non-invasive and has low power consumption. Doppler processing is an important part of many ultrasound imaging systems. It is used to provide blood velocity information and is built on top of B-mode systems. We investigate the performance of two velocity estimation schemes used in Doppler processing systems, namely, directional velocity estimation (DVE) and conventional velocity estimation (CVE). We find that DVE provides better estimation performance and is the only functioning method when the beam to flow angle is large. Unfortunately, DVE is computationally expensive and also requires divisions and square root operations that are hard to implement. We propose two approximation techniques to replace these computations. The simulation results on cyst images show that the proposed approximations do not affect the estimation performance. We also study backend processing which includes envelope detection, log compression and scan conversion. Three different envelope detection methods are compared. Among them, FIR based Hilbert Transform is considered the best choice when phase information is not needed, while quadrature demodulation is a better choice if phase information is necessary. Bilinear and Gaussian interpolation are considered for scan conversion. Through simulations of a cyst image, we show that bilinear interpolation provides comparable contrast-to-noise ratio (CNR) performance with Gaussian interpolation and has lower computational complexity. Thus, bilinear interpolation is chosen for our system.

Image understanding has been playing an increasingly crucial role in vision applications. Sparse models form an important component in image understanding, since the statistics of natural images reveal the presence of sparse structure. Sparse methods lead to parsimonious models, in addition to being efficient for large scale learning. In sparse modeling, data is represented as a sparse linear combination of atoms from a "dictionary" matrix. This dissertation focuses on understanding different aspects of sparse learning, thereby enhancing the use of sparse methods by incorporating tools from machine learning. With the growing need to adapt models for large scale data, it is important to design dictionaries that can model the entire data space and not just the samples considered. By exploiting the relation of dictionary learning to 1-D subspace clustering, a multilevel dictionary learning algorithm is developed, and it is shown to outperform conventional sparse models in compressed recovery, and image denoising. Theoretical aspects of learning such as algorithmic stability and generalization are considered, and ensemble learning is incorporated for effective large scale learning. In addition to building strategies for efficiently implementing 1-D subspace clustering, a discriminative clustering approach is designed to estimate the unknown mixing process in blind source separation. By exploiting the non-linear relation between the image descriptors, and allowing the use of multiple features, sparse methods can be made more effective in recognition problems. The idea of multiple kernel sparse representations is developed, and algorithms for learning dictionaries in the feature space are presented. Using object recognition experiments on standard datasets it is shown that the proposed approaches outperform other sparse coding-based recognition frameworks. Furthermore, a segmentation technique based on multiple kernel sparse representations is developed, and successfully applied for automated brain tumor identification. Using sparse codes to define the relation between data samples can lead to a more robust graph embedding for unsupervised clustering. By performing discriminative embedding using sparse coding-based graphs, an algorithm for measuring the glomerular number in kidney MRI images is developed. Finally, approaches to build dictionaries for local sparse coding of image descriptors are presented, and applied to object recognition and image retrieval.

Over the past fifty years, the development of sensors for biological applications has increased dramatically. This rapid growth can be attributed in part to the reduction in feature size, which the electronics industry has pioneered over the same period. The decrease in feature size has led to the production of microscale sensors that are used for sensing applications, ranging from whole-body monitoring down to molecular sensing. Unfortunately, sensors are often developed without regard to how they will be integrated into biological systems. The complexities of integration are underappreciated. Integration involves more than simply making electrical connections. Interfacing microscale sensors with biological environments requires numerous considerations with respect to the creation of compatible packaging, the management of biological reagents, and the act of combining technologies with different dimensions and material properties. Recent advances in microfluidics, especially the proliferation of soft lithography manufacturing methods, have established the groundwork for creating systems that may solve many of the problems inherent to sensor-fluidic interaction. The adaptation of microelectronics manufacturing methods, such as Complementary Metal-Oxide-Semiconductor (CMOS) and Microelectromechanical Systems (MEMS) processes, allows the creation of a complete biological sensing system with integrated sensors and readout circuits. Combining these technologies is an obstacle to forming complete sensor systems. This dissertation presents new approaches for the design, fabrication, and integration of microscale sensors and microelectronics with microfluidics. The work addresses specific challenges, such as combining commercial manufacturing processes into biological systems and developing microscale sensors in these processes. This work is exemplified through a feedback-controlled microfluidic pH system to demonstrate the integration capabilities of microscale sensors for autonomous microenvironment control.