Search results

Title: DEEP LEARNING FOR IMAGE PROCESSING WITH APPLICATIONS TO MEDICAL IMAGING
Creator: Zarshenas, Amin
Date: 2019
Description: Deep Learning is a subfield of machine learning concerned with algorithms that learn hierarchical data representations. Deep learning has...
Show moreDeep Learning is a subfield of machine learning concerned with algorithms that learn hierarchical data representations. Deep learning has proven extremely successful in many computer vision tasks including object detection and recognition. In this thesis, we aim to develop and design deep-learning models to better perform image processing and tackle three important problems: natural image denoising, computed tomography (CT) dose reduction, and bone suppression in chest radiography (“chest x-ray”: CXR). As the first contribution of this thesis, we aimed to answer to probably the most critical design questions, under the task of natural image denoising. To this end, we defined a class of deep learning models, called neural network convolution (NNC). We investigated several design modules for designing NNC for image processing. Based on our analysis, we design a deep residual NNC (R-NNC) for this task. One of the important challenges in image denoising regards to a scenario in which the images have varying noise levels. Our analysis showed that training a single R-NNC on images at multiple noise levels results in a network that cannot handle very high noise levels; and sometimes, it blurs the high-frequency information on less noisy areas. To address this problem, we designed and developed two new deep-learning structures, namely, noise-specific NNC (NS-NNC) and a DeepFloat model, for the task of image denoising at varying noise levels. Our models achieved the highest denoising performance comparing to the state-of-the-art techniques.As the second contribution of the thesis, we aimed to tackle the task of CT dose reduction by means of our NNC. Studies have shown that high dose of CT scans can increase the risk of radiation-induced cancer in patients dramatically; therefore, it is very important to reduce the radiation dose as much as possible. For this problem, we introduced a mixture of anatomy-specific (AS) NNC experts. The basic idea is to train multiple NNC models for different anatomic segments with different characteristics, and merge the predictions based on the segmentations. Our phantom and clinical analysis showed that more than 90% dose reduction would be achieved using our AS NNC model.We exploited our findings from image denoising and CT dose reduction, to tackle the challenging task of bone suppression in CXRs. Most lung nodules that are missed by radiologists as well as by computer-aided detection systems overlap with bones in CXRs. Our purpose was to develop an imaging system to virtually separate ribs and clavicles from lung nodules and soft-tissue in CXRs. To achieve this, we developed a mixture of anatomy-specific, orientation-frequency-specific (ASOFS) expert deep NNC model. While our model was able to decompose the CXRs, to achieve an even higher bone suppression performance, we employed our deep R-NNC for the bone suppression application. Our model was able to create bone and soft-tissue images from single CXRs, without requiring specialized equipment or increasing the radiation dose.
Show less

Title: Simulation, design and applications of a table top analyzer-based phase contrast mammography system
Creator: Caudevilla Torras, Oriol
Date: 2019
Description: Analyzer-based Imaging is a promising phase contrast technology with huge potential for soft tissue imaging. Unlike absorption-contrast...
Show moreAnalyzer-based Imaging is a promising phase contrast technology with huge potential for soft tissue imaging. Unlike absorption-contrast methods, phase-contrast modalities measure refraction and scatter properties of the tissue. Such images are particularly suitable for applications such as mammography.The potential advantages of the Analyzer-Based Imaging technology are three fold. First, it shows exceptional contrast when imaging soft tissue, which produces extremely sharp images of the breast compared to absorption images. Second, it provides additional insights about the breast. In particular, the density and scatter images of breast micro-calcifications can help assessing their malignancy better than common mammograms. Third, it has shown potential to reduce the radiation dose deposited in the breast tissue by an order of magnitude compared to common mammography procedures.In the past, Analyzer-Based Imaging has been mainly developed with synchrotron light sources and focused on obtaining micro-resolution images. For such applications, quasi-monoenergetic beams are required. Nevertheless, monochromatic radiation can be easily obtained in synchrotron setups by filtering the source’s spectrum with crystal optics. Since synchrotrons are very brilliant sources, most of their radiation can be filtered out and still obtain low noise phase contrast images. Nowadays, there is a lot of interest in transitioning the technology to a table-top system using compact X-ray sources for mammography. However, compact sources are several orders of magnitude less brilliant, which causes extremely long exposure times. Additionally, the trade-off between exposure time (throughput) and resolution in compact analyzer-based imaging systems is yet to be completely understood.In this thesis, we lay down the principles to develop compact analyzer-based imaging systems capable of imaging a full-sized breast under ten seconds, while ensuring a resolution under 100 microns. This represents a major breakthrough towards obtaining a clinical analyzer-based mammography system. Additionally, we explore a unique application of the analyzer-based technology for breast diagnosis consisting on the assessment of the chemical composition of micro-calcifications. In conjunction with ABI’s unparalleled image quality, determining the chemical composition of micro- calcifications can help to mitigate the high false positive rate in common mammography.
Show less

Title: Fast mesh based reconstruction for cardiac-gated SPECT and methodology for medical image quality assessment
Creator: Massanes Basi, Francesc
Date: 2018
Description: In this work, we are studying two different subjects that are intricately connected. For the first subject we are considering tools to...
Show moreIn this work, we are studying two different subjects that are intricately connected. For the first subject we are considering tools to improve the quality of single photon emission computed tomography (SPECT) imaging. Currently, SPECT images assist physicians to evaluate perfusion levels within the myocardium, aide in the diagnosis of various types of carcinomas, and measure pulmonary function. The SPECT technique relies on injecting a radioactive material into the patient's body and then detecting the emitted radiation by means of a gamma camera. However, the amount of radioactive material that can be given to a patient is limited by the negative effects that the radiation will have on the patient's health. This causes SPECT images to be highly corrupted by noise. We will focus our work on cardiac SPECT, which adds the challenge of the heart's continuous motion during the acquisition process. First, we describe the methodology used in SPECT imaging and reconstruction. Our methodology uses a content adaptive model, which uses more samples on the regions of the body that we want to be reconstructed more accurately and less in other areas. Then we describe our algorithm and our novel implementation that lets us use the content adaptive model to perform the reconstruction. In this work, we show that our implementation outperforms the reconstruction method used for clinical applications. In the second subject we are evaluating tools to measure image quality in the context of medical diagnosis. In signal processing, accuracy is typically measured as the amount of similarity between an original signal and its reconstruction. This similarity is traditionally a numeric metric that does not take into account the intended purpose of the reconstructed images. In the field of medical imaging, a reconstructed image is meant to aid a physician to perform a diagnostic task. Therefore, the quality of the reconstruction should be measured by how much it helps to perform the diagnostic task. A model observer is a computer tool that aims to mimic the performance of human observer, usually a radiologist, at a relevant diagnosis task. In this work we present our linear model observer designed to automatically select the features needed to model a human observer response. This is a novelty from the model observers currently being used in the medical imaging field, which instead usually have ad-hoc chosen features. Our model observer dependents only on the resolution of the image, not the type of imaging technique used to acquire the image.
Show less

Title: Intraoperative tumor margin detection using nanoparticles: protocol optimization through kinetic modeling
Creator: Xu, Xiaochun
Date: 2018
Description: Clear margins (no tumor on the surface of the resected tissues) is essential to minimize tumor recurrence and prolong survival for wide local...
Show moreClear margins (no tumor on the surface of the resected tissues) is essential to minimize tumor recurrence and prolong survival for wide local excision cancer surgeries. However, standard methods of margin assessment cannot be carried out within the time frame of surgery (meaning patients with positive margins are suggested to undergo call-back surgeries). Intraoperative molecular imaging of cell surface receptors can offer a solution; however, substantial nonspecific diffusion and retention of imaging agents in resected tissues remains a significant challenge to identifying cancer reliably. Recently, “paired-agent” methods—which employ co-administration of a control-imaging agent with a targeting agent—have been applied to thick-sample staining and rinsing applications to account for background staining. This dissertation aimed to optimize paired-agent molecular imaging tumor-to-healthy tissue discrimination through mathematical modeling.Two simplified mathematical models—the rinsing paired-agent model (RPAM) and the serial staining model (SSM)—were derived and tested in accurate simulation models (also developed as a component of this dissertation,) and in preclinical cancer models. More specifically, RPAM was demonstrated to be capable of providing more accurate estimates of receptor concentration than more standard “ratiometric” methods (essentially dividing the targeted agent signal by the control agent signal), and the model was insensitive to the variability of rinsing time from one image to the next. Though it was noted in experiments, that regardless of the approach taken, a very large fraction of signal was removed upon the first rinse, leading to large “gaps” in the data that would be available to RPAM. The SSM, on the other hand, provided a model that could be applied to serial staining data, which yielded a more gradual change in signal between imaging.Considering the multidimensional complexity of paired-agent topical tissue molecular imaging (with diffusion, imaging agent chemical/binding properties, tissue staining, rinsing, imaging, and data analysis protocols all being subject to alteration), thorough optimization margin analysis imaging protocols is untractable using experiments alone. Therefore, a salient feature of this dissertation was the development and validation of a “forward” mathematical diffusion and binding model for in silico testing of proposed paired-agent staining and rinsing protocols in thick tissue.
Show less

Title: STRATEGIES TO MAXIMIZE DOSE REDUCTION IN SPECT MYOCARDIAL PERFUSION IMAGING
Creator: Juan Ramon, Albert
Date: 2019
Description: Radiation exposure in medical imaging has become a topic of major concern, gaining intense attention within the clinical and research...
Show moreRadiation exposure in medical imaging has become a topic of major concern, gaining intense attention within the clinical and research communities. In 2009, the National Council on Radiation Protection and Measurements (NCRP) announced radiation exposure of patients via medical imaging increased more than sixfold between the 1980s and 2006, with cardiac nuclear medicine, specifically myocardial perfusion imaging (MPI) with single-photon emission computed tomography (SPECT) being the second biggest culprit. The goal of this work is to evaluate several strategies to enable radiation dose to be minimized while maintaining current levels of diagnostic accuracy in the clinic. We achieve dose reduction through optimization of advanced image reconstruction strategies, to obtain higher-quality images at a given dose (noise) level, through a machine learning approach to predict the optimal dose for each patient, and through advanced deep learning (DL) algorithms to improve the quality of reconstructed images. Our ultimate objective is to provide the nuclear cardiology field with a new set of algorithms and guidelines for selecting administered activity levels and image reconstruction procedures in the clinic. The project is based on a clinical study in which imaging and various other data are being collected for a set of patients. The project has the following components. First, we investigate a global dose-reduction approach (i.e., reducing dose by a uniform proportion across all patients) via optimization of image reconstruction strategies. Specifically, we maximize perfusion-defect detection (diagnostic accuracy) over a range of simulated dose levels using clinical data into which we have introduced simulated defects. We measure diagnostic performance using clinically validated model observers from the Quantitative Perfusion SPECT (QPS) software package. We investigate the diagnostic accuracy over a range of dose levels ranging from those currently used in the clinic down to one-eighth of this level. We consider the following image-reconstruction: filtered-backprojection (FBP) with no correction for physics effects, and ordered-subsets expectation-maximization (OS-EM) with several combinations of attenuation correction (AC), scatter correction (SC), and resolution correction (RC).Second, we propose a patient-specific ("personalized") dose reduction approach based on machine learning that aims to predict the minimum radiation dose needed to obtain consistent perfusion-defect detection accuracy for each individual patient. This prediction is based on patient attributes, especially body measurements, and various clinical variables. We compare the diagnostic accuracy produced by predicted personalized doses to that produced by standard clinical dose levels to validate the predictive models.Third, we verify that the dose minimization results obtained in the context of perfusion-defect detection also maintain diagnostic accuracy in evaluating cardiac function, as characterized by myocardial motion.Finally, we propose a deep learning (DL) method to denoise SPECT-MPI reconstructed images. The method is a 3D convolutional neural network trained to predict standard-dose images from low-dose images. We quantify the extent to which dose reduction can be achieved using the proposed DL structure when dose is reduced uniformly across patients or by means of our patient-specific approach.
Show less

Title: Quantification of Vascular Permeability in the Retina Using Fluorescein Videoangiography Data as a Biomarker for Early Diabetic Retinopathy
Creator: Kayaalp Nalbant, Elif
Date: 2023
Description: Diabetic retinopathy, which is the most common reason for blindness in the working-age population, affects over one-third of those who have...
Show moreDiabetic retinopathy, which is the most common reason for blindness in the working-age population, affects over one-third of those who have had diabetes for over ten years. High blood sugar level (hyperglycemia) in the blood damages blood vessels and tight junction at the blood-retinal barrier (BRB). Chronic inflammation leads to changes in vascular health, and over time blood vessels tend to get damaged and exhibit higher “leakage” or permeability. In the late stage of DR, hemorrhages can occur, leading to irreversible damage of neuronal tissue in the retina and vision loss. In the clinic, there are some biomarkers and imaging modalities used to diagnose DR based on some of the more severe products of DR (e.g., hemorrhage), but there is no non-invasive, highly sensitive method to detect diabetic retinopathy before clinical signs occur, when mitigating therapies could be more effective. In this thesis, indicator dilution theory was explored to modeling the temporal dynamics of fluorescein in the retina after intravenous injection, with an aim to quantitatively map subtle changes in retinal blood flow and vascular permeability that could preempt subsequent irreversible damage. Specifically, a simplified version of indicator dilution theory—namely the “adiabatic approximation in tissue homogeneity” (AATH) model—was used to estimate physiological parameters such as the blood flow (F) and the extraction fraction (E: a parameter coupled with vascular permeability) from retinal fluorescein videoangiography data. The AATH fitting protocol was optimized through simulations using a more complex model (the AATH-vascular heterogeneity model, AATH-VH). It was determined that a two-step least square fitting method was more sensitive than a single-step least square fitting of AATH to simulated data to evaluate vascular permeability in early diabetic retinopathy. The optimized data analysis protocol was then evaluated in an initial clinical study comparing healthy control subjects to those with moderate non-proliferative DR. Volumetric blood flow and retinal vascular permeability maps were compared between patient groups with clear increases in extraction fraction observed in the mild NPDR patients compared to control. These promising early data have been the foundation to an ongoing 5 year study tracking 100 Diabetic patients with no DR so see if early changes in vascular permeability can predict which patients are more likely to progress to DR.
Show less

Title: DEVELOPMENT AND EVALUATION OF MRI TEMPLATES OF THE MIITRA ATLAS
Creator: RIDWAN, ABDUR RAQUIB
Date: 2021
Description: Digital human brain atlases play a pivotal role in conducting wide range of neuroimaging studies and are commonly used as references for...
Show moreDigital human brain atlases play a pivotal role in conducting wide range of neuroimaging studies and are commonly used as references for spatial normalization in voxel-wise analysis, region-of interest analyses, automated tissue-segmentation, functional connectivity analyses, etc. A brain atlas typically consists of MRI-based multi-modal templates and semantic labels delineating brain regions according to the characteristics of the underlying tissue. In recent times there has been a plethora of magnetic resonance imaging (MRI) studies on older adults without dementia to explore the role of brain characteristics associated with cognitive functions in old age with the ultimate goal to develop strategies for prevention of cognitive decline. Increasing the accuracy in terms of sensitivity and specificity of such neuroimaging studies require an atlas with a comprehensive set of high-quality templates representative of the brain characteristics typical of older adults and detailed labels accurately mapping brain regions of interest. However, such an atlas has not been constructed for older adults without dementia. Hence this thesis aims to build high quality MRI templates which are the cornerstone resources needed for the development of a comprehensive, high quality, multi-channel, longitudinal, probabilistic digital human brain atlas for older adults termed as Multi-channel Illinois Institute of Technology and Rush University Aging (MIITRA) atlas. This dissertation focuses on a) to develop and evaluate a high performing 1mm isotropic structural T1-weighted brain template, b) to investigate the development and evaluation of a spatio-temporally consistent longitudinal structural T1-weighted template of the older adult brain, c) to develop and evaluate an unbiased 0.5 mm isotropic super-resolved high resolution and detail-preserving structural T1 weighted template of the older adult brain, d) to develop an unbiased 0.5 mm super-resolved high resolution and detail-preserving structural PD weighted and T2-weighted template of the older adult brain, e) to investigate and provide future directions in the development of a 0.5 mm super-resolved high resolution DTI template of the older adult brain, and f) to construct a novel approach in the development of MRI templates using both space and frequency information of spatially normalized older adult data. The thesis based on the aforementioned foundational points was constructed as follows: Firstly, this thesis presents the development of a 1mm isotropic T1-weighted structural template of the older adult brain utilizing state of the art registration algorithm ANTs with parameters carefully optimized for older adults, in an iterative groupwise spatial normalization framework. The preprocessing steps were also thoroughly investigated to ensure high quality data. It was demonstrated through systematic comparison of this new template to several other standardized and study-specific T1-weighted templates that a) it exhibited high image sharpness, b) allowed for high spatial normalization accuracy and detection of smaller inter-group morphometric differences compared to other standardized templates, c) had similar performance to that of study-specific templates and d) was highly representative of the older adult brain. Secondly, with the acquired technical know-how from the aforementioned research findings a new method was introduced for the construction of a spatio-temporally consistent longitudinal template based on high quality cross-sectional older adult data from a large cohort. The new template was compared to templates generated with previously published methods in terms of spatio-temporal consistency and image quality and was shown to have superior performance. In addition, a novel approach was introduced for image quality enhancement of the longitudinal templates utilizing both space and frequency information. Thirdly, the thesis presents a method that involves a) thoroughly refining registration parameters, b) patch-based tissue-guided sparse-representation approach in a super-resolved unbiased minimum deformation space to construct and evaluate an unbiased 0.5 mm isotropic super-resolved high resolution and detail-preserving structural T1 weighted template of the older adult brain. This method accounts for misregistration specially in the cortical regions, ensuring sharp delineation of structures representative of the older adult brain. The new template developed using this approach maintained high anatomical consistency with sharp and detailed cortical features in the brain and exhibited higher image sharpness compared to other high-resolution standardized templates and allowed for high spatial normalization accuracy when used as a reference for normalization of older adult data. Additionally, this approach of template building was investigated on DTI tensors of older adult participants, and the constructed DTI template was shown to perform better than templates developed using the best approach currently present in the literature. Finally, the thesis presents the development of an unbiased 0.5 mm super-resolved high resolution and detail-preserving structural PD weighted and T2-weighted template of the older adult brain, from nonlocal super-resolution based upsampled PD and T2w older adult participant data, using this new template building approach.
Show less

Title: Development and evaluation of high resolution MRI templates and labels of the MIITRA atlas
Creator: Niaz, Mohammad Rakeen
Date: 2022
Description: A digital human brain atlas consisting of MRI-based multi-modal templates and semantic labels delineating brain regions are commonly used as...
Show moreA digital human brain atlas consisting of MRI-based multi-modal templates and semantic labels delineating brain regions are commonly used as references for spatial normalization in a wide range of neuroimaging studies. Magnetic resonance imaging (MRI) studies of the aging brain is of significant interest in recent times to explore the role of brain characteristics associated with cognitive functions. The introduction of advanced image reconstruction techniques, and the recent trend in MRI acquisitions at submillimeter in-plane resolution have resulted in an easier availability of MRI data on older adults at high spatial resolution. An atlas with a comprehensive set of high-resolution templates representative of the older adult brain and detailed labels accurately mapping brain regions can increase the sensitivity and specificity of such neuroimaging studies. Additionally, most neuroimaging studies can benefit from a high-resolution atlas with templates where fine brain structures are resolved and, where the transition between different tissue can be more accurately defined. However, such an atlas is not publicly available for older adults. Hence the goal of this thesis is to develop a comprehensive, high-resolution digital human brain atlas for older adults termed as Multi-channel Illinois Institute of Technology and Rush University Aging (MIITRA) atlas.This dissertation aims a) to develop a new technique based on the principles of super-resolution for the construction of high-resolution structural and diffusion tensor templates, and evaluate the templates for use in studies on older adults, b) to construct and evaluate high-resolution structural and diffusion tensor templates constructed using the method developed in (a) for the MIITRA atlas using MRI data collected on 400 nondemented older adults, c) to investigate and develop a technique for the construction of high-resolution labels and evaluate the performance of gray matter labels constructed using this technique in segmenting the gray matter of older adults, and d) to develop and evaluate a comprehensive set of high-resolution labels using the technique developed in (c) for the MIITRA atlas using data on 400 non-demented older adults. Based on the aforementioned points, the thesis is structured as follows: Firstly, this thesis presents a novel approach for the construction of a high-resolution T1-weighted structural template based on the principles of super resolution. This method introduced a forward mapping technique to minimize signal interpolation, and a weighted averaging method to account for residual misregistration. The new template was shown to resolve finer brain structures compared to a lower resolution template constructed using the same data. It was demonstrated through systematic comparison of this new template to several other standardized templates of different resolutions that a) it exhibited high image sharpness, b) was free of image artifacts, c) allowed for high spatial normalization accuracy and detection of smaller inter-group morphometric differences compared to other standardized templates, d) was highly representative of the older adult brain. This novel approach was further modified for the construction of a high spatial resolution diffusion tensor imaging template. The new DTI template is the first high spatial resolution population-based DTI template of the older adult brain and exhibits high image quality, high sharpness, is free of artifacts, resolves fine white matter structures, and provides higher spatial normalization accuracy of older adult DTI data compared to other available DTI templates. Secondly, the aforementioned techniques were utilized in the development of high resolution T1-weighted and DTI templates, and tissue probability maps for the MIITRA atlas using high quality MRI images on 400 diverse, community cohort of non-demented older adults. Thirdly, a novel approach for generating high resolution gray matter labels is presented that involves a) utilization of the super resolution technique to ensure sharp delineation of structures, and b) a multi atlas based correction technique to reduce errors due to misregistration. High-resolution gray matter labels were constructed using the super resolution technique. When used for regional segmentation of the gray matter of older adults, the new gray matter labels of the showed high overlap, high geometric correlation, and low dissimilarity with the manually edited reference labels, demonstrating that there is a high agreement between the new labels and the manually edited Freesurfer labels. Finally, this thesis presents the development of a comprehensive array of gyral-based, cytoarchitecture-based, and functional connectivity-based gray matter labels in MIITRA space utilizing the aforementioned techniques. These labels include gyral-based, cytoarchitecture-based, and functional connectivity-based labels which will enhance the functionality of the MIITRA atlas. The new labels will also enhance the interoperability of MIITRA with the source atlases.
Show less

repository.iit

Search the repository

Enabled Filters

Refine Results

Date

Subject

Creator