Browsing by Author "Souza, Roberto"

Now showing 1 - 7 of 7
  • Thumbnail Image
    ItemOpen Access
    A voxel-level approach to brain age prediction: A quantitative method to assess regional brain aging
    (2023-12-05) Gianchandani, Neha; Souza, Roberto; MacDonald, Ethan; Bayat, Sayeh; Pike, Bruce; Harris, Ashley; Tan, Benjamin
    Global brain age has been used as an effective biomarker to study the correlation between brain aging and neurological disorders. However, brain aging is a regional phenomenon, a facet that remains relatively under-explored within the realm of brain age prediction research using machine learning methods. Voxel-level predictions can provide localized brain age estimates that can provide granular insights into the regional aging processes. This is essential to understand the differences in aging trajectories in healthy versus diseased subjects. In this work, a deep learning- based multitask model is proposed for voxel-level brain age prediction. The proposed model outperforms the model existing in the literature and yields valuable clinical insights when applied to both healthy and diseased populations. Most findings from the analysis align with existing studies on aging, whereas other findings are intriguing and could be potential biomarkers of early-stage neurodegeneration detection. Regional analysis is performed on the voxel-level brain age predictions to understand aging trajectories of known anatomical regions in the brain and show that there exist disparities in regional aging trajectories of healthy subjects compared to ones with underlying neurological disorders such as dementia and more specifically, Alzheimer’s disease. A comparative analysis with traditional deep learning interpretability methods showed that the proposed voxel-level approach to brain age prediction is an effective way to understand regional aging trajectories while being quantitative in nature. The source code is publicly available at https://github.com/nehagianchandani/Voxel-level-brain-age-prediction.
  • Thumbnail Image
    ItemEmbargo
    AI-Assisted Interactive Assistants for Software Issue Report Understanding
    (2024-06-04) Tamanna, Salma Begum; Uddin, Gias; Souza, Roberto; Abdellatif, Ahmad
    Issue reports in software projects often become complex due to their technical details and lengthy discussions, leading to information overload. This complexity can hinder quick understanding of these reports, impacting the development process adversely. This thesis investigates whether automatic assistance can help tackle the problem. It first introduces the iSum (issue summarizer) tool, designed to generate visual summaries of information types present in issue reports and analyze the prevalence and trends of these across a report or a repository. Next, it addresses a RAG-based ChatGPT’s struggle with understanding complex technical content from bug reports and interpreting context from queries for exploring bug reports. Our enhancement, the ChatGPT Inaccuracy Mitigation Engine (CHIME), boosts response correctness of ChatGPT by around 30%. Both iSum and CHIME demonstrate the potential of AI to enhance the comprehensibility of issue reports, taking a step forward in efficient issue understanding.
  • Thumbnail Image
    ItemOpen Access
    An Integrated Deep Learning Model with Genetic Algorithm (GA) for Optimal Syngas Production Using Dry Reforming of Methane (DRM)
    (2024-01-17) Zarabian, Maryam; Olatunji Fapojuwo, Abraham; Souza, Roberto; Clarke, Matthew Alexander
    The dry reforming of methane is a chemical process transforming two primary sources of greenhouse gases, i.e., carbon dioxide (CO2) and methane (CH4), into syngas, a versatile precursor in the industry, which has gained significant attention over the past decades. Nonetheless, commercial development of this eco-friendly process faces barriers such as catalyst deactivation and high energy demand. Artificial intelligence (AI), specifically deep learning, accelerates the development of this process by providing advanced analytics. However, deep learning requires substantial training samples and collecting data on a bench scale encounters cost and physical constraints. This study fills this research gap by employing a pretraining approach, which is invaluable for small datasets. It introduces a software sensor for regression (SSR) powered by deep learning to estimate the quality parameters of the process. Moreover, combining the SSR with a genetic algorithm offers a prescriptive analysis, suggesting optimal thermodynamic parameters to improve the process efficiency.
  • Thumbnail Image
    ItemOpen Access
    Analysis of Deep Domain Adaptation Methods for Brain Magnetic Resonance Image Segmentation
    (2022-12-16) Saat, Parisa; Hemmati, Hadi; Souza, Roberto; Gavrilova, Marina; Deshpande, Gouri
    Accurate brain segmentation is critical for magnetic resonance imaging (MRI) analysis pipelines. Machine-learning-based brain MR image segmentation methods are among the state-of-the-art techniques for this task. Nevertheless, the segmentations produced by machine learning models often degrade in the presence of expected domain shifts between the test and train sets data distributions. These domain shifts are expected due to several factors, such as scanner hardware and software differences, technology updates, MRI acquisition parameters, and differences across the scanned populations. Domain adaptation (DA) methods can make machine learning models more resilient to these domain shifts. In this thesis, I investigated supervised and unsupervised deep domain adaptation methods for brain MRI segmentation. Two scenarios are investigated. In the first scenario, data shifts occur due to hardware and software differences across different MRI scanner vendors (General Electric, Philips, and Siemens). In the second scenario, data shifts occur due to differences in the scanned populations. The source brain MRI data comes from adults, while the target data corresponds to pediatric patients, whose brains are still developing. The main findings of this thesis are that there is not a single DA technique that consistently outperforms others, and hyperparameter tuning and computational times for these methods still need to be addressed before adopting these methods in clinical practice. Another important finding is that the DA techniques worked better for data shifts resulting from hardware and software differences across different MR scanner vendors than data shifts from population differences. The labeled data and source code used in this thesis were made publicly available and serve as a benchmark for evaluating DA methods for brain MRI segmentation.
  • Thumbnail Image
    ItemOpen Access
    Analysis of Stroke Induced Motor Function Weakness in Post-Stroke Patients using Machine Learning
    (2021-09) Bhatt, Aakash; Yanushkevich, Svetlana; Souza, Roberto; Nielsen, John
    The focus of this thesis is the development of a system that can analyze stroke induced motor weakness using pressure sensor mattresses. The proposed system utilizes time-series pressure data from publicly available datasets as well as data collected from recovering stroke patients at the Foothills Medical Center. Two tasks are performed with the pressure data. In the first task the incoming pressure data is classified into three body positions: supine, lateral, and prone on a frame-by-frame basis. The second task consists of classifying time-series pressure data into two classes: left-sided weakness and right-sided weakness. Results from the first task are used to improve results from the second task by only using patient data in which the patient is in a supine position. Extensive experiments are conducted using deep learning methodologies including convolutional neural networks and long-short term memory networks. The developed system is intended to be utilized to monitor patient condition throughout their stay at the hospital.
  • Thumbnail Image
    ItemOpen Access
    Brain connectomes in youth at risk for serious mental illness: an exploratory analysis
    (2022-09-15) Metzak, Paul D.; Shakeel, Mohammed K.; Long, Xiangyu; Lasby, Mike; Souza, Roberto; Bray, Signe; Goldstein, Benjamin I.; MacQueen, Glenda; Wang, JianLi; Kennedy, Sidney H.; Addington, Jean; Lebel, Catherine
    Abstract Background Identifying early biomarkers of serious mental illness (SMI)—such as changes in brain structure and function—can aid in early diagnosis and treatment. Whole brain structural and functional connectomes were investigated in youth at risk for SMI. Methods Participants were classified as healthy controls (HC; n = 33), familial risk for serious mental illness (stage 0; n = 31), mild symptoms (stage 1a; n = 37), attenuated syndromes (stage 1b; n = 61), or discrete disorder (transition; n = 9) based on clinical assessments. Imaging data was collected from two sites. Graph-theory based analysis was performed on the connectivity matrix constructed from whole-brain white matter fibers derived from constrained spherical deconvolution of the diffusion tensor imaging (DTI) scans, and from the correlations between brain regions measured with resting state functional magnetic resonance imaging (fMRI) data. Results Linear mixed effects analysis and analysis of covariance revealed no significant differences between groups in global or nodal metrics after correction for multiple comparisons. A follow up machine learning analysis broadly supported the findings. Several non-overlapping frontal and temporal network differences were identified in the structural and functional connectomes before corrections. Conclusions Results suggest significant brain connectome changes in youth at transdiagnostic risk may not be evident before illness onset.
  • Thumbnail Image
    ItemOpen Access
    Transferring Transfer Function (TTF): A Guided Approach to Transfer Function Optimization in Volume Visualization
    (2024-04-22) Nasim Saravi, Amin; Alim, Usman; Alim, Usman; Costa Sousa, Mario; Souza, Roberto
    In volume visualization, a transfer function tailored for one volume usually does not work for other similar volumes without careful tuning. This process can be tedious and time-consuming for a large set of volumes. Leveraging the differentiability of volume rendering, this work presents a novel approach to transfer function optimization based on a reference volume and its transfer function. Using two fully connected neural networks, our approach learns a continuous 2D separable scalar-gradient transfer function that visualizes the features of interest with consistent visual properties across volumes. The resulting optimized transfer function is exportable for further interactions in domain-specific applications. Through two compelling use cases—tracking the aftermath of an asteroid blast near the ocean surface and visualizing brain white matter, grey matter, and cerebral fluid in magnetic resonance (MR) images—we demonstrate the effectiveness of our approach, validated through collaboration with domain experts.