<p dir="ltr">Functional magnetic resonance imaging (fMRI) and diffusion tensor imaging (DTI) via MRI are powerful, noninvasive methods for imaging of the human brain. Here, two studies are presented which explore algorithm development for the processing and analysis of fMRI and DTI-MRI data.</p><p dir="ltr">In the first study, brain functional connectivity was analyzed in a cohort of high school American football athletes over a single play season and compared against participants in non-collision high school sports. Football athletes underwent four resting-state functional magnetic resonance imaging sessions: once before (pre-season), twice during (in-season), and once 34–80 days after the contact activities play season ended (post-season). For each imaging session, functional connectomes (FCs) were computed for each athlete and compared across sessions using a metric reflecting the (self) similarity between two FCs. HAEs were monitored during all practices and games throughout the season using head-mounted sensors. Relative to the pre-season scan session, football athletes exhibited decreased FC self-similarity at the later in-season session, with apparent recovery of self-similarity by the time of the post-season session. In addition, both within and post-season self-similarity was correlated with cumulative exposure to head acceleration events. These results suggest that repetitive exposure to HAEs produces alterations in functional brain connectivity and highlight the necessity of collision-free recovery periods for football athletes.</p><p dir="ltr">In the second study, a method for harmonization of DTI-MRI data across sites was assessed. Pooling of data from multiple sites is limited by noise characteristics of individual scanners and their receive chain elements (e.g., coils, filters, algorithms), requiring careful consideration of methods to harmonize multisite data. Here, the ComBat data harmonization method was assessed on DTI-MRI data to determine if the harmonizing transformation produced by the algorithm could be transferred to harmonize new subject data from previously-observed sites without necessitating reharmonization of pre-existing data. Results indicated that this transferable ComBat methodology (T-ComBat) yielded reduced differences in fractional anisotropy and mean diffusivity across sites when compared with unharmonized data but did not fully reach the performance of ComBat applied to the entire dataset. Results of this study provide guidelines for circumstances (namely, the proportion of subjects one may wish to add to an existing dataset) under which T-ComBat may be effectively applied to harmonize new subject DTI-MRI data.</p>

<p dir="ltr">Fluorescence microscopy, a cornerstone in biological imaging, faces inherent challenges due to photon budget constraints that affect the signal-to-noise ratio (SNR), ultimately limiting imaging performance. This thesis explores theoretical frameworks to address two fundamental issues: the denoising limit of fluorescence microscopy images and the resolution limit in the presence of photon noise. Firstly, we extend the application of the Cramér-Rao Lower Bound (CRLB) to establish a variance lower bound for image denoising algorithms in fluorescence microscopy. By incorporating constraints specific to the imaging system and biological specimens, we provide a benchmark for evaluating the performance of state-of-the-art denoising algorithms. Our analysis reveals that this lower bound is determined by factors such as photon count, readout noise, detection wavelength, effective pixel size, and numerical aperture of the microscope system. Secondly, building upon the pioneering work by Ernest Abbe and leveraging modern fluorescence and nanoscopy advancements, we propose a novel theoretical framework to quantify the resolving power of fluorescence microscopes under finite photon conditions. This model integrates the traditional diffraction limit with photon statistics to determine the practical resolution limit, highlighting the trade-offs between photon noise and resolution enhancement in techniques like confocal microscopy. This dual approach not only refined the theoretical understanding of fluorescence microscopy's capabilities but also assisted in designing and optimizing more effective imaging protocols. Through these investigations, this thesis provided a comprehensive theoretical foundation for improving fluorescence microscopy imaging techniques, paving the way for future innovations in biological imaging.</p>

Computed tomography (CT) combined with deep learning (DL) has recently shown great potential in biomedical imaging. Complex DL models with varying architectures inspired by the human brain are improving imaging software and aiding diagnosis. However, the accuracy of these DL models heavily relies on the datasets used for training, which often contain low-quality CT images from low-dose CT (LDCT) scans. Moreover, in contrast to the neural architecture of the human brain, DL models today are dense and complex, resulting in a significant computational footprint. Therefore, in this work, we propose sparse optimizations to minimize the complexity of the DL models and leverage architecture-aware optimization to reduce the total training time of these DL models. To that end, we leverage a DL model called DenseNet and Deconvolution Network (DDNet). The model enhances LDCT chest images into high-quality (HQ) ones but requires many hours to train. To further improve the quality of final HQ images, we first modified DDNet's architecture with a more robust multi-level VGG (ML-VGG) loss function to achieve state-of-the-art CT image enhancement. However, improving the loss function results in increased computational cost. Hence, we introduce sparse optimizations to reduce the complexity of the improved DL model and then propose architecture-aware optimizations to efficiently utilize the underlying computing hardware to reduce the overall training time. Finally, we evaluate our techniques for performance and accuracy using state-of-the-art hardware resources. / Master of Science / Deep learning-based (DL) techniques that leverage computed tomography (CT) are becoming omnipresent in diagnosing diseases and abnormalities associated with different parts of the human body. However, their diagnostic accuracy is directly proportional to the quality of the CT images used in training the DL models, which is majorly governed by the radiation dose of the X-ray in the CT scanner. To improve the quality of low-dose CT (LDCT) images, DL-based techniques show promising improvements. However, these techniques require substantial computational resources and time to train the DL models. Therefore, in this work, we incorporate algorithmic techniques inspired by sparse neural architecture of the human brain to reduce the complexity of such DL models. To that end, we leverage a DL model called DenseNet and Deconvolution Network (DDNet) that enhances the quality of CT images generated by low X-ray dosage into high-quality CT images. However, due to its architecture, it takes hours to train DDNet on state-of-the-art hardware resources. Hence, in this work, we propose techniques that efficiently utilize the hardware resources and reduce the time required to train DDNet. We evaluate the efficacy of our techniques on modern supercomputers in terms of speed and accuracy.

<p dir="ltr">High-frequency ultrasound is an excellent non-invasive imaging technique that we utilized in both the quantification of intervertebral disc strain and microrobots for targeted drug delivery systems due to its ability to provide real-time, high-resolution images of soft tissues without the need for invasive procedures. In the intervertebral disc study, it allows us to capture detailed strain measurements across time during dynamic loading, offering insights into tissue deformation under physiological conditions. Similarly, in the microrobot study, high-frequency ultrasound enables precise visualization and tracking of the microrobots within the tissue environment, ensuring accurate navigation and monitoring during targeted drug delivery.</p><p><br></p><p dir="ltr">Understanding the mechanics of intervertebral disc injury and validating computational models requires accurate measurement of internal disc strain. While noninvasive imaging and advanced image processing have enabled strain quantification, these methods often depend on visual markers that may alter tissue behavior. Furthermore, they are typically limited to static testing, which does not fully capture physiological loading conditions. This study aims to utilize high-frequency ultrasound combined with texture correlation techniques to quantify strain within intervertebral discs during dynamic axial loading. This study demonstrates that high-frequency ultrasound is an effective tool for measuring disc strain under dynamic loading conditions.</p><p><br></p><p dir="ltr">Inflammatory bowel disease (IBD) refers to a group of disorders characterized by chronic inflammation of the digestive tract. Common treatments for IBD, such as oral and injectable medications, are selected based on disease severity but present significant challenges. These include systemic toxicity, non-specific drug targeting, and degradation of the therapeutic agents in the upper gastrointestinal tract, which reduces treatment effectiveness before reaching the colon. The goal of this research is to develop and test microrobots for targeted drug delivery to inflamed regions of the colon in IBD patients. The microrobots are guided by external magnetic torque and aim to overcome the limitations of traditional IBD treatments by delivering smaller, more effective drug payloads directly to the site of inflammation. The study demonstrates that microrobots, guided by magnetic torque, can effectively deliver targeted drug payloads to inflamed regions of the colon. This method allows for a smaller drug payload compared to traditional treatments, offering a more precise and efficient approach for treating IBD.</p>

Acenes have been explored by a number of research groups in the field of organic electronics with a particular emphasis on transistor materials. This group has been actively studying acene‐based organic semiconductors for more than a decade using a crystal engineering approach and has developed acene derivatives for applications in field‐effect transistors, light‐emitting diodes, and photovoltaics. In addition to organic electronics, crystal engineering has important applications in a number of other fields, quite notably in the design of metal‐organic frameworks. Chapters 2 and 3 of this dissertation focus on applying crystal engineering to the synthesis of acene derivatives for use as solid‐state, long‐wavelength fluorescent organic dyes in the field of biomedical imaging. More specifically, this work studied the synthesis and properties of dioxolane‐functionalized pentacenes and hexacenes. One of these pentacene derivatives has already been demonstrated in biomedical imaging which may lead to improved treatment of tuberculosis. The dioxolane‐functionalized hexacene is still under evaluation for bioimaging applications. Chapters 4 and 5 focus on crystal engineering in relation to organic electronics. Chapter 4 deals with fine‐tuning of crystal packing and demonstrated that small differences in molecular structure can result in significant changes to the solid‐state structure which affects semiconductor properties. Finally, chapter 5 studies the use of singlet fission in photovoltaics and demonstrated that this process does occur in a solar cell incorporating a hexacene derivative. Pentadithiophenes were also synthesized for singlet fission photovoltaics, but they have yet to be studied further.

Chemical exchange saturation transfer (CEST) MRI has the ability to noninvasively measure endogenous biomarkers and exogenous agents relevant to various diseases and medical conditions. My work has focused on the development of MRI pulse sequences and data analysis methods to more accurately estimate endogenous and exogenous CEST contrast measurements at 7 T and 3 T magnetic field strengths. Chapter 1 discusses the various sources of signal that have been measured with CEST MRI in the clinic, the acquisition methods used to acquire these signals, and the data analysis methods used to quantify the CEST effects from these signals. Appendix A describes the development of a respiration gated CEST pulse sequence that was ultimately used with a lung fibrosis mouse model to measure extracellular pH (pHe) of the fibrotic lesions. Appendix B describes the development of a data processing algorithm that used the Bloch equations modified for chemical exchange to generate more accurate and precise pHe estimates both at 7 T and 3 T magnetic field strengths relative to a previous data processing algorithm. Appendix C describes the development of a retrospective gating technique for the lung that generates more accurate and precise endogenous CEST contrast measurements.

Technological advances in next-generation sequencing and biomedical imaging have led to a rapid increase in biomedical data dimension and acquisition rate, which is challenging the conventional data analysis strategies. Modern machine learning techniques promise to leverage large data sets for finding hidden patterns within them, and for making accurate predictions. This dissertation aims to design novel machine learning-based models to transform biomedical big data into valuable biological insights. The research presented in this dissertation focuses on three bioinformatics domains: splice junction classification, gene regulatory network reconstruction, and lesion detection in mammograms. A critical step in defining gene structures and mRNA transcript variants is to accurately identify splice junctions. In the first work, we built the first deep learning-based splice junction classifier, DeepSplice. It outperforms the state-of-the-art classification tools in terms of both classification accuracy and computational efficiency. To uncover transcription factors governing metabolic reprogramming in non-small-cell lung cancer patients, we developed TFmeta, a machine learning approach to reconstruct relationships between transcription factors and their target genes in the second work. Our approach achieves the best performance on benchmark data sets. In the third work, we designed deep learning-based architectures to perform lesion detection in both 2D and 3D whole mammogram images.

Laparoscopic surgery or minimally invasive surgery has great advantages compared with the conventional open surgery, such as reduced pain, shorter recovery time and lower infection rate. It has become a standard clinical procedure for cholecystectomy, appendectomy and splenectomy. The state-of-the-art laparoscopic technologies suffer from several significant limitations, one of which is the tradeoff of the limited instantaneous field of view (FOV) for high spatial resolution versus the wide FOV for situational awareness but with diminished spatial resolution. Standard laparoscopes lack the ability to acquire both wide-angle and high-resolution images simultaneously through a single scope. During the surgery, a trained assistant is required to manipulate the laparoscope. The practice of frequently maneuvering the laparoscope by a trained assistant can lead to poor or awkward ergonomic scenarios. This type of ergonomic conflicts imposes inherent challenges to laparoscopic procedures, and it is further aggravated with the introduction of single port access (SPA) techniques to laparoscopic surgery. SPA uses one combined surgical port for all instruments instead of using multiple ports in the abdominal wall. The grouping of ports raises a number of challenges, including the tunnel vision due to the in-line arrangement of instruments, poor triangulation of instruments, and the instrument collision due to the close proximity to other surgical devices. A multi-resolution foveated laparoscope (MRFL) was proposed to address those limitations of the current laparoscopic surgery. The MRFL is able to simultaneously capture a wide-angle view for situational awareness and a high-resolution zoomed-in view for fine details. The high-resolution view can be scanned and registered anywhere within the wide-angle view, enabled by a 2D optical scanning mechanism. In addition, the high-resolution probe has optical zoom and autofocus capabilities, so that the field coverage can be dynamically varied while keep the same focus distance as the wide-angle probe. Moreover, the MRFL has a large working distance compared with the standard laparoscopes, the wide-angle probe has more than 8x field coverage than a standard laparoscope. On the other hand, the high-resolution probe has 3x spatial resolution than a standard one. These versatile capabilities are anticipated to have significant impacts on the diagnostic, clinical and technical aspects of minimally invasive surgery. In this dissertation, the development of the multi-resolution foveated laparoscope was discussed in detail. Starting from the refinement of the 1st order specifications, system configurations, and initial prototype demonstration, a customized dual-view MRFL system with fixed optical magnifications was developed and demonstrated. After the in-vivo test of the first generation prototype of the MRFL, further improvement was made on the high-resolution probe by adding an optical zoom and auto-focusing capability. The optical design, implementation and experimental validation of the MRFL prototypes were presented and discussed in detail.

Biomechanical imaging (BMI) is the process of non-invasively measuring the spatial distribution of mechanical properties of biological tissues. The most common approach uses ultrasound to non-invasively measure soft tissue deformations. The measured deformations are then used in an inverse problem to infer local tissue mechanical properties. Thus quantifying local tissue mechanical properties can enable better medical diagnosis, treatment, and understanding of various diseases. A major difficulty with ultrasound biomechanical imaging is getting accurate measurements of all components of the tissue displacement vector field. One component of the displacement field, that parallel to the direction of sound propagation, is typically measured accurately and precisely; the others are available at such low precision that they may be disregarded in the first instance. If all components were available at high precision, the inverse problem for mechanical properties could be solved directly, and very efficiently. When only one component is available, the inverse problem solution is necessarily iterative, and relatively speaking, computationally inefficient. The goal of this thesis, therefore, is to develop a processing method that can be used to recover the missing displacement data with sufficient precision to allow the direct reconstruction of the linear elastic modulus distribution in tissue. This goal was achieved by using a novel spatial regularization to adaptively enforce and locally relax a special form of momentum conservation on the measured deformation field. The new processing method was implemented with the Finite Element Method (FEM). The processing method was tested with simulated data, measured data from a tissue mimicking phantom, and in-vivo clinical data of breast masses, and in all cases it was able to recover precise estimates the full 2D displacement and strain fields. The recovered strains were then used to calculate the material property distribution directly.

Parmi les approches d'imagerie préclinique, les techniques optiques sur petit animal fournissent une information fonctionnelle sur un phénomène biologique ainsi que sur sa localisation. De récents développements permettent d'exploiter ces méthodes dans le cadre de l'imagerie sur animal vigile. Les conditions physiologiques se rapprochent alors de celles du fonctionnement normal de l'organisme. Les travaux de cette thèse ont porté sur l'utilisation optimale de cette modalité via des méthodes originales d'analyse et de traitement.Les problèmes soulevés par la fusion des flux cinématiques et de données de bioluminescence nous ont amené à proposer des approches complémentaires d’estimationde mouvement de l’animal. La représentation sous forme implicite des informations issuesde la vidéo de l’animal permettent de construire un critère robuste à minimiser. L’ajout d’uncritère global mesurant la compacité du signal optique permet de considérer dans sa totalité les données multicanaux acquises pour augmenter la précision du recalage. Finalement ces deux modélisations offrent des résultats pertinents et validés expérimentalement.Dans le but de s'affranchir des contraintes de l'observation planaire de nos données nous avons conçu une méthode d’estimation du mouvement 3D de l’animal à partir d’un modèle pré-calculé. Grâce à un système d'acquisition multi-vues et simultanée de la scène, il est possible d’ajouter une contrainte sur l'estimation de la position de la source pour rendre robuste le suivi des poses issues de la vidéo. Les résultats expérimentaux montrent le potentiel de cette méthode pour fournir des mesures 3D précises sur l'animal vigile. / Optical imaging techniques, have taken, since many years, a great part in the preclinicalstudies. The luminescence signal could be now recorded with a short time resolution whichenables studies with freely moving animals. This is an improvement because several studieshighlighted the impact of anesthetics agent and animal handling to perform studies inphysiological conditions. In this thesis, we define the tools, based on computer visionmethods, which offer the possibility to express the potential of this modality.In some cases, animal movement and low signal produce weak localization of the signal.Therefore we propose to improve localization of the optical data for a freely moving animal byusing motion field obtained from the multi-channel data. First, we introduce silhouetteconstraints and landmarks on the mouse skin within a variation framework. To take intoaccount all data in the registration framework, we combine the previously defined criteria,with global ones which measure compactness of signal distribution. Fusion is formulated as adiscrete population framework which produces strong experimental results in comparison topairwise method.In the last part, we propose an original approach to enable 3D optical imaging in case offreely moving animal. Therefore, we present a novel model-based method to animal trackingfrom monocular video which allows the 3D measurement of the signal. The 3D animal poseand the illumination are dynamically estimated through minimization of an objective functionwith constraints on the signal position. Experimental results demonstrate the potential of ourapproach for 3D accurate measurement with freely moving animal.