22689 research outputs found
Sort by
Towards Interpretable Machine Learning in Medical Image Analysis
No full textOver the past few years, ML has demonstrated human expert level performance in many medical image analysis tasks. However, due to the black-box nature of classic deep ML models, translating these models from the bench to the bedside to support the corresponding stakeholders in the desired tasks brings substantial challenges. One solution is interpretable ML, which attempts to reveal the working mechanisms of complex models. From a human-centered design perspective, interpretability is not a property of the ML model but an affordance, i.e., a relationship between algorithm and user. Thus, prototyping and user evaluations are critical to attaining solutions that afford interpretability. Following human-centered design principles in highly specialized and high stakes domains, such as medical image analysis, is challenging due to the limited access to end users. This dilemma is further exacerbated by the high knowledge imbalance between ML designers and end users. To overcome the predicament, we first define 4 levels of clinical evidence that can be used to justify the interpretability to design ML models. We state that designing ML models with 2 levels of clinical evidence: 1) commonly used clinical evidence, such as clinical guidelines, and 2) iteratively developed clinical evidence with end users are more likely to design models that are indeed interpretable to end users. In this dissertation, we first address how to design interpretable ML in medical image analysis that affords interpretability with these two different levels of clinical evidence. We further highly recommend formative user research as the first step of the interpretable model design to understand user needs and domain requirements. We also indicate the importance of empirical user evaluation to support transparent ML design choices to facilitate the adoption of human-centered design principles. All these aspects in this dissertation increase the likelihood that the algorithms afford interpretability and enable stakeholders to capitalize on the benefits of interpretable ML. In detail, we first propose neural symbolic reasoning to implement public clinical evidence into the designed models for various routinely performed clinical tasks. We utilize the routinely applied clinical taxonomy for abnormality classification in chest x-rays. We also establish a spleen injury grading system by strictly following the clinical guidelines for symbolic reasoning with the detected and segmented salient clinical features.
Then, we propose the entire interpretable pipeline for UM prognostication with cytopathology images. We first perform formative user research and found that pathologists believe cell composition is informative for UM prognostication. Thus, we build a model to analyze cell composition directly. Finally, we conduct a comprehensive user study to assess the human factors of human-machine teaming with the designed model, e.g., whether the proposed model indeed affords interpretability to pathologists. The human-centered design process is proven to be truly interpretable to pathologists for UM prognostication. All in all, this dissertation introduces a comprehensive human-centered design for interpretable ML solutions in medical image analysis that affords interpretability to end users
MECHANO-INDUCED SELF-DRIVEN, REVERSIBLE, HOMOTYPIC MONOCYTE DOMAIN FORMATION
No full textIt has been well-documented that matrix stiffness plays an indispensable role in cell phenotype and function. Before the transition into tissue-resident macrophages, monocytes patrol inside blood vessels. Upon sensing the signals of infection or tissue damage, monocytes can extravasate and invade surrounding tissue. The elastic modulus of surrounding tissue varies depending on its location and function. The stiffness value of healthy tissue is significantly smaller than that of diseased tissue. For example, the stiffness of lung solid tumors (20-30kPa) is stiffer than that of normal lung parenchyma (0.5-5kPa) (1). However, how mechanical signals influence monocyte behaviors is still under-investigated. Herein, we used polyacrylamide hydrogels of 0.5kPa, 100kPa, and normal tissue culture plastic to represent the stiffnesses of normal tissue, diseased tissue, and common cell culture condition, respectively. We found that mediated by the intracellular mechano-sensing pathways, monocytes show different domain formations and development processes on the three substrates. We identified β2 integrin as an indispensable mediator of the aggregation process by applying both inhibitory and stimulatory antibodies on monocytes. Furthermore, we proposed the “Local Activation, Global Inhibition” theory combining the local activation of β2 integrin and global inhibitory effects of self-secreted, diffusible molecules. In addition, we performed a series of experiments to determine possible parameters such as seeding density which may influence the domain formation process. A preliminary computational model was developed to validate our findings. In this study, we describe mechano-induced monocyte homotypic domain formation and the underlying “Local Activation, Global Inhibition” mechanism. Collectively, this study suggests that stiffness is important in modulating immune cell behavior and function. In vitro culture conditions of immune cells should be carefully selected to avoid huge discrepancies in their in vivo phenotypes and behaviors. Leveraging substrate stiffness to modulate immune response may be a possible therapeutic strategy in the future
USING CULTURALLY RESPONSIVE TEACHING PRACTICES TO INCREASE TEACHERS’ SELF-EFFICACY WHEN TEACHING DIVERSE LEARNERS
No full textMany urban public school teachers are under prepared to teach culturally and linguistically diverse students. Often, there is a disconnect between non-White students from diverse cultural backgrounds and urban school teachers, many of whom are White. This cultural mismatch between students and teachers can negatively affect student learning and expose some monolingual teachers’ limited personal experiences and knowledge of culturally responsive teaching practices (CRTP) when teaching culturally and linguistically diverse students. The researcher conducted a mixed-method study to examine the impact of such cultural disconnect on instruction in a diversely populated urban public school, using professional development and coaching interventions. Researchers have suggested that teachers’ cultural competency and sensitivity toward culturally diverse learners are linked to students’ overall academic achievement. Teachers who are culturally responsive understand cultural differences and similarities amongst cultural groups and races and embed students’ diverse cultural backgrounds as strengths in daily instructional practices
A QUANTITATIVE ANALYSIS OF ADAPTATIONS BY PERINATAL PSYCHIATRY ACCESS PROGRAMS TO PROMOTE MENTAL HEALTHCARE EQUITY
No full textPerinatal Psychiatry Access Programs (“access programs”) build the capacity of healthcare professionals in the perinatal field. As a systems-level intervention, the access programs aim to improve perinatal mental healthcare. This study uses multivariate regression analysis to examine (1) state and program characteristics associated with mental healthcare equity adaptations and (2) state factors associated with the creation of access programs. The analysis showed that both state and program characteristics correlated with mental healthcare equity adaptations. The state qualities that showed a positive relationship with creating an access program were federal funding, the legislative majority party, the percent of births to marginalized people, and median household income. These findings reveal that decisions regarding equity adaptations can lead to more targeted interventions within the national network of access programs as well as provide insight. Understanding the critical factors associated with states could inform the existing network to ensure success in creating new programs
EXPANDING ENERGY ACCESS THROUGH MINIGRIDS IN SUB-SAHARAN AFRICA PRIVATE SECTOR, FINANCING APPROACHES, AND PUBLIC – PRIVATE MODELS: WHAT WORKS?
No full textA study on the financial viability of electricity utilities in Africa (M. C. Kojima 2016, 1-37) found that of 39 countries, only the Seychelles and Uganda fully recovered their operational and capital costs. In addition, only 19 countries had utilities that collected sufficient cash to cover operating costs, with just four recovering half or more of the capital costs when new replacement values of current assets were considered. The findings showed that such significant funding gaps prevent power sectors from delivering reliable electricity to existing customers, challenging supply extensions to new consumers at an optimal pace. Moreover, due to their weak financial situation, African energy utility companies have no or low borrowing capacity, and the governments have been assuming debt for capital expenditure (CAPEX) needs. Because of this, utilities cannot (i) make the capital investments needed to expand access to new consumers or (ii) maintain existing assets, which leads to low or poor quality of service for electricity consumers. The shortfall in financing in the energy sector in many Sub-Saharan African (SSA) countries stems from various factors, including excess power generation capacity, electricity tariffs that are not cost-reflective (almost in all countries), and significant technical and commercial and collection losses in distribution. In addition to the lack of financing for the supply and distribution of electricity, affordability by the end user is another major impediment to electricity access in SSA, compounded by poverty, low and inconsistent income, high cost of imported fuel for power generation in some countries, structural inequalities, unwillingness to pay for services by consumers, and high connection fees. This makes private-sector financing extremely important in the quest of governments to reach more people with affordable electricity access. While minigrids are part of the solution to scale up energy access, the optimal technology for achieving universal electrification in SSA consists of grid, minigrid, and Solar Home Systems (SHS). Countries with a comprehensive approach involving this combination of solution(s) have achieved the fastest results in electricity access (Energy Sector Management Assistance Program 2019, 5).
Minigrids are suitable for rural electrification when they have the lowest cost (unsubsidized electricity retail cost on site in USD/kWh) compared to grid extension and stand-alone systems over the medium to long-term period (i.e., 10 to 30 years) (Energy Sector Management Assistance Program 2019). Minigrids can be deployed quickly under the right conditions, which include the availability of financing and appropriate regulatory frameworks. If host governments allow cost-recovery tariffs, minigrids can be profitable and attractive to investors and financiers while providing critical social services required to improve SSA's human capital and social development (Ehrhardt 2019, 1-7). Through this research, the nexus between (i)models of financing for minigrids, (ii) minigrid ownership models, (iii) increase in electricity access, (iv) replicability, (v) scalability and (vi) sustainability will be analyzed. The research methodology consists of (i) a literature review, (ii) expert interviews, (iii) secondary data analysis, and iv) case studies of selected countries in Sub-Saharan Africa. The findings will contribute to increasing electricity access in SSA through minigrids, providing information to guide policymakers and stakeholders and motivating private investors to invest more in minigrids in Sub-Saharan Africa.
Primary Reader and Advisor: Sarah Jordaan, Ph.D.
Secondary Reader: Matthew Kocher, Ph.D.
Third Reader: Koffi Ekouevi, Ph.D
Investigating Polycarbonate Welds made with Fused Filament Fabrication
No full textFused Filament Fabrication (FFF) offers low cost, rapid production times,
and greater design flexibility than traditional manufacturing methods like injection molding. The primary drawback of FFF parts is their high variability and reduced mechanical, and fracture properties relative to injection molded parts. While the
greater portion of the reduction and variability in properties are attributed to defects from printing, FFF samples without defects have been shown to preferentially fracture at the welds.
To understand why failure localizes to the welds, and how the intrinsic
strength of the welds develops, this dissertation combines optical characterization of the local geometry of the weld with in-situ infrared measurement of the weld
thermal history to underpin mechanical characterization methods to directly
investigate the fracture and failure of individual welds.
The work undertaken in this dissertation develops novel tests to probe and investigate the intrinsic strength of inter-road polymer welds. Three tests are discussed, a trouser tear test, a mixed-mode peel test, and a razorblade-induced fracture test. The trouser tear test is rapidly becoming a standard research tool to probe welds. However, the sample size and cross-section shape were found to significantly impact the measured tear strength of a weld, even after post-print heat treatment to overwrite the printing thermal history. The mixed-mode peel test successively peels individual filaments to measure the fracture energy of the weld as a function of printing temperature and layer height, two commonly varied printing parameters. The fracture energy increases with the weld time, from 0.1
N/mm to 0.8 N/mm, to a maximum of 80% of a possible bulk value in less than a second of welding.
The last test developed is a razorblade-induced fracture test with an incremental method that measures the sample-blade friction during the retraction of the blade. The measured fracture energy for injection-molded sheets was found to be sample geometry independent and the fracture energy of the weld reaches 70% of that of injection molded sheets (31 N/mm). However, cutting welds shows fast fracture rather than a highly plastic cutting that is seen in the injection-molded sheets and a reduction in the plastic zone around the crack tip. The shape of the cross-section introduces a change in failure behavior, and a change in the plastic zone around the crack tip as seen through ex-situ birefringence. The consistent
findings have been that the shape of the cross-sections can cause a stress
concentration, making the welds appear weaker than bulk. The tests explored in this work provide new tools and guidance to explore the quality of polymer welds produced by FFF or other manufacturing methods
DESIGN OF A WHOLE-GENOME CRISPR SCREEN IN ER+ HER2-MUTANT BREAST CANCER AND INVESTIGATING ESR1 NOVEL AND TRANSACTIVATING MUTATIONS
No full textEstrogen receptor-positive (ER+) breast cancer is a common oncological malignancy accounting for roughly 60-70% of breast cancers. In early stages of these cancers, the current treatment modalities are ER-targeted therapies such as aromatase inhibitors, selective estrogen receptor modulators (SERMs) and selective estrogen receptor degraders (SERDs). However, due to cancer evolution and therapeutic selective pressure, endocrine resistant tumors are common in the later stages of ER+ breast cancer. Common mechanisms of resistance include acquired mutations in ER (ESR1), HER2 (ERBB2), or other genes involved in the mitogen-activated protein kinase (MAPK) pathway. While the resistance mechanisms of ESR1 mutations are well studied, the mechanism whereby HER2 mutations confer resistance remains unknown. In this research, we performed two projects to investigate the mechanisms of resistance of both the ER and HER2 receptor. To investigate HER2 mutations, we optimized conditions to perform a whole-genome CRISPR screen to determine the specific pathways that allow ER+ HER2 mutant tumors to be re-sensitized to standard treatments. In the second project, we focused on investigating the effects of novel single and double ESR1 mutations in response to standard of care therapies within the clinic. Understanding the molecular mechanisms that mutations within these two key receptors implicated in breast cancer has significant clinical implications to allow researchers to develop alternative or combination treatments for patients
Geographic Context and Early Childhood Educator Perceptions: Associations with Job Demands and Resources, Classroom Quality, and Child Outcomes in Head Start Settings
No full textHead Start teachers play a crucial role in promoting children’s development of social, emotional, and behavioral skills, particularly through consistent, emotionally supportive interactions (Whitaker et al., 2015). The conditions of teachers’ work, including the demands of their role and the resources they are able to access, are associated with teachers’ physical (Bronkhorst, 2015) and psychological well-being (e.g., stress, burnout, depressive symptoms) (Skaalvik & Skaalvik, 2018; Zinsser et al., 2019; Kwon et al., 2020). As such, job demands could limit their ability to establish a positive classroom environment that facilitates positive growth and development in children (Jennings & Greenberg, 2009). Across geographic contexts (i.e., region, urban density, poverty density), school climate, resources, economic landscape, and socialization of children may vary greatly (Conger, 2013). Given the importance of context for teachers’ perceptions and child development (Bronfenbrenner, 1976), I examine associations between geographic context, resources and demands, classroom quality, and social, emotional, and behavioral outcomes (approaches to learning, social skills, and behavioral functioning) in children.
To address dearth of research on these associations in the early care and education (ECE) literature, this dissertation uses secondary data from the U.S. Department of Agriculture (USDA), the 2014-2017 cohort of the Family and Child Experiences Survey (FACES), and the 2019 National Survey of Early Care and Education (NSECE). The USDA provides data on child poverty across the nation, as well as region and urbanicity. FACES captures data about classroom activities, classroom observations, demographic characteristics, family engagement, policies (reported by center directors, education coordinators), and child reports (reported by teachers). NSECE is unique in its consideration of resources and demands of the work environment for ECE staff, regional context (west, south, midwest, northeast), urban density, and poverty density (low, moderate, high) in understanding teachers’ and children’s experiences in early care and education. The goals of this project are (a) to describe the landscape of quality of life for young children and families in the United States (U.S.), (b) examine the distribution of ECE workforce job demands and resources across geographic context, (c) to understand the role of staff job demands and resources in the association between geographic context and teacher beliefs (i.e., caregiving beliefs, teacher-child interactions), and (d) to analyze direct and indirect associations between teacher collaborative practices and children’s outcomes via classroom quality
Acute Watery Diarrhea Related Healthcare Seeking Behavior in Chittagong Bangladesh
No full textCholera surveillance in Bangladesh relies on a passive, clinical system that fails to cover most of the country’s population. By only counting those that present for care at designated surveillance facilities the system may miss various demographic and spatial groups of people. While these data are vital to estimating the burden of disease in Bangladesh, current modeling approaches do not account for care seeking trends that otherwise could be used to increase the accuracy of burden estimates. This analysis used hypothetical, report, and clinical care seeking data collected from Sitakunda Upazila, Bangladesh in 2021 and 2022 to explore care seeking trends in the study population with the goal of informing future burden of disease calculations. Generalized estimating equations and generalized linear models were used by the analysis to assess the strength of association of various independent variables. Analysis of these data found that as severity of hypothetical acute watery diarrhea (AWD, used as a proxy measurement for suspected cholera) increased, respondents were more willing to indicate that they would seek care. Where individuals would seek care was also impacted by the level of AWD severity, having implications on how researchers should control for variations in healthcare seeking trends when dealing with clinical data. Comparing hypothetical to reported care seeking, a notable qualitative shift across the null in the association of SES, household income, household distance to port, and household distance to the nearest health center was observed. We call for further studies to be conducted that are better powered to observe care seeking associations among individuals that reported experiencing AWD. These efforts could improve disease burden modeling practices in the future
Using a Single Neural Network for Simultaneous Noise Reduction and Layer Segmentation in Visible Light Optical Coherence Tomography of human retina
No full textVisible light optical coherence tomography (VIS-OCT) is an emerging imaging modality that uses shorter wavelength in visible light range than conventional near infrared (NIR) light. It provides one-micron level axial resolution to improve image contrast to better separate stratified retinal layers, as well as provides microvascular oximetry with spatio-spectral analysis. However, due to the practical limitation of laser safety and comfort, the permissible illumination power is much lower than NIR OCT which can be challenging to obtain high quality VIS-OCT images and subsequent image analysis particularly in pathological eyes. Therefore, improving VIS-OCT image quality by denoising is an essential step in the overall workflow in VIS-OCT clinical applications. In this thesis, we provide the first VIS-OCT retinal image dataset from normal eyes, including retinal layer annotation and noisy-clean image pairs. We propose an efficient co-learning deep learning framework for noisy-input segmentation embedded with both supervised and self-supervised denoising process, and with personalized percentage of segmentation labels to alleviate the laborious annotation work. The same neural network performed both denoising and segmentation tasks simultaneously. The task performance is benchmarked qualitatively and quantitatively. The significant improvement of segmentation (2% higher Dice coefficient compared to segmentation-only process) for GCL, IPL, INL is observed when available annotation drops to 25%, indicating a potential angle for annotation-efficient training. We also showed that the denoising model trained on our dataset generalizes well to a human retina volume obtained from a different scanning protocol