Browsing by Subject "Deep Learning"
Now showing 1 - 8 of 8
- Results Per Page
- Sort Options
Item Addressing Label Variability in Deep Learning-Based Segmentation for Radiation Therapy(December 2021) Balagopal, Anjali; Wang, Jing; Jiang, Steve B.; Nguyen, Dan; Lin, Mu-Han; Fei, BaoweiAccurate segmentation of tumor and surrounding organs-at-risk (OARs) is important for radiotherapy treatment(RT) planning. Manual segmentation by physicians currently used in clinical practice is time consuming and highly depends on the physicians' skill and experience, leading to large inter-and intra-observer variation. Deep learning(DL) algorithms, such as those used for image recognition, has been promising in the development of automated segmentation tools for medical imaging. Contouring in RT has some specific challenges as opposed to semantic segmentation in other spaces. DL segmentation models are trained with large, annotated datasets. The annotations play such an important role since these labels are the supervisory signals teaching the model how to segment. Majority of the times, a very high-quality dataset is unavailable. The datasets available are usually biased, noisy, and sometimes scarcely annotated. One of the primary sources of noise in RT datasets is annotation variations. This dissertation addresses the challenge of label variability and investigates different methods to deal with various kinds of label variability. The thesis studies the impact that label variability has on automation performance as well as patient outcome and devices ways to reduce or to respect this variability. Computed Tomography based segmentation model for intact and post-operative prostate cancer RT planning are proposed and developed for clinical use. Uncertainty in the automated label prediction for CTV and OARs are evaluated in detail. Impact of inter-physician variability on patient outcome for post-operative prostate cancer is investigated and instead of segmenting a general CTV , PSA-Net is proposed that respects style variations across physicians as well as across institutions. For dealing with systematic label variability in some structures, a prior-guided deep difference meta-learner is proposed that can segment a structure in a new labelling style from just a handful of prior segmented patients. A multi-modality IPA segmentation model is proposed to reduce label variability due to expertise differences among physicians in clinical trials. This model can effectively help inexperienced physicians in producing expert segmentations. The contributions of this thesis are expected to facilitate better understanding of label variability in RT and help in avoiding/respecting label variability when developing deep learning models for RT.Item Automated Treatment Planning in High Dose-Rate Brachytherapy for Cervical Cancer(2021-05-01T05:00:00.000Z) Gonzalez, Yesenia Amanda; Wang, Jing; Medin, Paul; Hrycushko, Brian; Albuquerque, KevinStandard care of cervical cancer is chemoradiation therapy followed by a boost to the cervix tumor site with High-dose-rate brachytherapy (HDRBT). HDRBT is a procedure that involves the insertion of radioactive sources into the tumorous area to ablate the cancer. Treatment planning for the procedure is typically performed on the day of treatment. The complex treatment planning process has led to issues such as prolonged treatment planning time and suboptimal plan quality. This dissertation reports systematic studies to improve treatment planning for HDRBT of cervical cancer. After a brief overview in Chapter 1, Chapter 2 will analyze the problems of sub-optimal plan quality and time management in the treatment planning workflow at our institution. In Chapter 3, developments on computational modules for automatic organ segmentation will be presented. Current clinical practice primarily relies on manual organ segmentation, but we have developed deep-learning models to segment the bladder, rectum, and sigmoid colon automatically, organs of particular importance for cervical cancer HDRBT. In Chapter 4, I present integration of the organ segmentation modules and other modules into the AutoBrachy system for clinical use to automate the planning process. Chapter 5 presents a deep-learning based method that can predict the physician's preference for a patient-specific treatment plan as defined by EQD2 to the bladder, rectum, sigmoid, and CTV D90. This method will serve as a guide in the future to automatically create patient-specific physician preferred treatment plans. Chapter 6 reports the study analyzing benefits from the use of joint intracavitary and interstitial HDRBT. Finally, Chapter 7 concludes the dissertation with discussions and future work.Item The Capabilities of Neural Systems Depend on a Hierarchically Structured World(August 2021) Blazek, Paul Joseph; Reynolds, Kimberly A.; Pfeiffer, Brad E.; Toprak, Erdal; Zinn, Andrew R.; Lin, MiloThe study of the human mind spans thousands of years, from the earliest philosophers to modern neuroscience. However, there still remains an incredible gap in our understanding of how cognitive functions arise from the biology of the brain. This has greatly hampered attempts to understand how the brain works, what its limitations are, and how to replicate it in artificial intelligence systems. Here I propose a general framework to understand how cognitive processes can be encoded by networks of interconnected neurons. I have taken a theoretical and computational approach, using artificial neural networks as a high-level quantitative model of basic neuroscientific principles. Neural networks are capable of reasoning by means of a series of specialized distinctions made by individual neurons that are integrated hierarchically. This framework enables the study of how the capabilities of neural systems are dependent on structural and functional constraints. Biological constraints on neural coding and network size and topology limit the complexity of stimuli that can be comprehended by the network. Surprisingly, though, neural networks are capable of comprehending much more complex stimuli than has been previously described, provided that the inherent distinctions between these stimuli are hierarchically structured. Functional constraints on neural networks require that they be able to perform cognitive processes and be able to reason in a way that can be communicated with other people. I have proposed a novel neurocognitive model which, when implemented in deep neural networks, is able to simulate a wide variety of cognitive processes. It is consistent with experimental evidence from neuroscience and theories from philosophy and psychology. By directly implementing symbolic reasoning within the structure and function of the network, it becomes possible to overcome many of the fundamental problems that face modern artificial intelligence systems, including their lack of explainability, robustness, and generalizability. This culminated in a novel algorithm that translates neural networks to human-understandable code, providing a complete picture of how neural networks can reason. All of these results suggest that neural systems require the world to be hierarchically structured in order to comprehend it, a direct reflection of their own hierarchical organization.Item Comprehensive Analysis of Lung Cancer Prognostic Factors(2019-07-29) Wang, Shidan; Gerber, David E.; Xie, Yang; Xiao, Guanghua; Zhan, Xiaowei; Hoshida, YujinLung cancer is the leading cause of death from cancer. It is remarkably heterogeneous in histopathological features and highly variable in prognosis. Analysis of prognostic factor is anticipated to guide clinicians for treatment selection, enhance patient care, and help understanding biological mechanism of tumor progression. To extend current knowledge about lung cancer prognosis, this dissertation analyzed lung cancer prognostic factors in three levels. First, in tumor level, deep learning aided pathology image analysis was used to extract tumor geometry and microenvironment features, upon which an image-based survival prediction model was built and independently validated for lung adenocarcinoma. Second, in patient level, a nomogram was built with demographic and clinical variables for patients with small cell lung cancer. The nomogram was implemented online for public usage. Third, in population level, how facility type and volume affect survival outcome and surgery selection for early stage non-small cell lung cancer was analyzed.Item Correcting Biases in Multi-Module Neural Networks, Through Efficient Hyperparameter Optimization and Statistically Meaningful Uncertainty Quantification, with Applications to Neurological Disorders(2023-05-01T05:00:00.000Z) Treacher, Alex H.; Lin, Milo; Rajaram, Satwik; Jaqaman, Khuloud; Vinogradov, Elena; Montillo, Albert A.Deep learning is a branch of machine learning that employs artificial neural networks to produce inferences from data. These networks have been successfully applied to a plethora of clinically and biologically related problems, including prognosis and diagnosis for a broad spectrum of diseases. However, many applications have focused on single tasks and/or single modal data, and thus use networks consisting of a single module. Modules are sections of a complete network, each that carries out a specific task. Multi-module artificial neural networks take inspiration from mammalian brains that process different types of input via dedicated brain regions (e.g. optical information by the visual cortex, sound, and language via Broca's and Wernicke's areas) and subsequently integrates them into a unified representation of our surroundings. While much work has been done on networks with a single module, significantly less work has focused on multi-module networks. Such networks are especially important in clinical problems that require the integration of multi-modal information and/or extracting multiple representations from the same input to provide high predictive performance. This dissertation corrects biases in multi-module networks for clinically relevant neurological problems. It develops novel hyperparameter search strategies with significantly improved performance of multi-module networks for diagnoses and prognoses, and adds uncertainty quantification to empower multi-module mixed effects deep learning models with the ability to produce statistically meaningful measures of covariate significance and principled probabilistic prediction confidence. Concretely, in this dissertation, I demonstrate the ability of multi-module deep learning networks to integrate spatial and temporal information and automate the detection of artifacts in magnetoencephalography (MEG) brain recordings of subjects including control subjects and those with a head injury. Recognizing how important the optimization of the network architecture was to achieve high predictive performance motivated the development of a novel module adaptive hyperparameter optimization (MA) hyperparameter search framework, which increases the efficiency of architecture optimization of multi-module networks. This approach is demonstrated to identify more optimal architectures when compared to other search strategies and to significantly increase the predictive performance of Alzheimer's and Parkinson's disease prognoses. Finally, I empower mixed effects deep learning (MEDL) models, which explicitly use multi-module networks, with uncertainty quantification allowing for the calculation of fundamental statistical metrics of model fit, covariate coefficient estimation, and prediction confidence. This model is then applied to predict which subjects will convert from mild cognitive impairment to Alzheimer's disease.Item Development of Deep Learning Artificial Intelligence to Detect Osteoporosis(2024-01-30) Fan, Christopher; Scanio, Angelo; Joshi, Parag; Öz, Orhan K.; Peshock, Ronald M.; Kay, FernandoOsteoporosis poses a substantial social and economic burden, with estimated treatment costs reaching a combined six trillion USD in the USA, Canada, and Europe. Although dual-energy X-ray absorptiometry (DEXA) is the diagnostic gold standard, computed tomography (CT) scans have proven to be reliable proxies for bone density measurement. Opportunistic screening for low bone density using CT obtained for other purposes can potentially reduce complications from osteoporotic fractures and health care costs. In this study, we developed an artificial intelligence (AI) algorithm using neural networks and the MONAI library to estimate DEXA bone density from non-contrast cardiac CT obtained for coronary calcium scoring purposes. A total of 2797 Dallas Heart Study phase 2 participants (39% male, 61% female) were included. The AI algorithm was first developed to automatically segment trabecular bone from cortical bone. This was trained and validated with manual segmentation of the trabecular bone by two medical students, a radiologist, using MONAI 3D autoseg. The ML algorithm achieved a Dice score of 0.97 when compared to human segmentation. A second AI model was developed utilizing segmentations of the first model. This AI was trained utilizing corresponding DEXA bone mineral density (BMD) for thoracic vertebrae. The best performing model was trained for 102 epochs, resulting in a training root mean square error (RMSE) of 0.0628 mg/cm2 and validation RMSE of 0.0842 mg/cm2. The final AI algorithm predictions yielded an R2 value of 0.71 compared to DEXA (Figure 1). Our findings underscore the clinical feasibility for an automated neural network to predict DEXA scores from non-contrast cardiac CT. This approach may help in the early detection of unsuspected low bone mineral density in patients undergoing CT scans for other reasons, allowing for potential improvements in patient outcomes and resourceful utilization of diagnostic imaging.Item Efficient and Intelligent Radiotherapy Planning and Adaptation(2022-05) Ma, Lin; Gu, Xuejun; Lu, Weiguo; Jiang, Steve B.; Jia, Xun; Wang, JingTreatment planning--inverse planning on volumetric tomography images--is the foundation of modern radiotherapy. Contouring planning structures and optimizing plan dose distributions are the two most important components of treatment planning. Treatment planning happens in two stages. The initial treatment plan is first designed before the commencement of treatment course. With the on-line images acquired before each fraction of treatment, treatment plan may be adapted off-line/on-line accounting for inter/intra-fractional changes. This dissertation presents innovative research to improve the efficiency and intelligence of treatment planning, with a special focus on its application in on-line adaptive planning. Four standalone, but related, techniques are developed in this dissertation. * Registration-guided on-line image segmentation. This technique targets the segmentation of on-line images. Based on the previous methods for on-line image contouring such as registration-based contour adaptation and deep learning-based image segmentation, we proposed a technique that can combine the merits of individual method. * Volumetric dose extension and isodose tuning. Interactive plan dose tuning is essential to on-line plan adaptation. We developed a dose painting algorithm that can output volumetric dose distribution from two isodose surfaces in real time. Then we applied the algorithm to interactive plan dose tuning, which allows for tuning volumetric dose by dragging isodose lines. * Fluence map prediction. Plan optimization is the central part of inverse planning. We proposed a deep learning-based fluence map prediction method, to achieve inverse planning without optimization. Compared to fluence map optimization, fluence map prediction is as accurate and much faster because of its nature as a direct inference calculation. * On-line proton range verification. Proton therapy is sensitive to motion and anatomy variation. For the previously proposed range-guided proton therapy strategy, we conducted an end-to-end Monte-Carlo simulation of the on-line proton measurement process. The key parameters of range measurement, such as the mapping from measurement result to range and measurement uncertainty, are obtained by simulation. Clinical data has been used to train and evaluate each technique. The results show the feasibility, generalizability, as well as limitations of the developed techniques. Future directions for on-line planning were discussed.Item Improving Cone-Beam Computed Tomography Based Adaptive Radiation Therapy with Deep Learning(2023-05-01T05:00:00.000Z) Liang, Xiao; Nguyen, Dan; Jiang, Steve B.; Lin, Mu-Han; Wang, Jing; Lu, WeiguoDuring radiation therapy, patient anatomical changes may compromise treatment quality if the treatment plan, prepared prior to therapy, remains unchanged throughout the course. Adaptive radiation therapy (ART) has been developed to address this issue by adapting the treatment plan based on up-to-date patient anatomy. The widespread availability of cone-beam computed tomography (CBCT) and its capability for the 3D imaging of patient anatomy has made CBCT-based ART an emerging and increasingly popular technology in the field of radiation oncology. However, numerous challenges persist in the current workflow. One challenge involves generating synthetic CT (sCT) images that retain CBCT anatomy while maintaining CT image quality. Clinically used sCT is typically obtained through deformable image registration (DIR) between pre-planning CT (pCT) and CBCT; however, this method often inadequately preserves CBCT anatomy due to DIR errors. Another challenge stems from the truncation issue in CBCT images, resulting from size limitations of imaging panels, which consequently leads to inaccuracies in CBCT-based dose calculations. Additionally, auto-segmentation of CBCT is impeded by low image quality and a lack of training labels. Although addressing this issue is difficult, enhancing auto-segmentation is essential, as manual segmentation is highly time-consuming. This dissertation aims to improve CBCT-based ART by leveraging deep learning (DL) technologies. First, an unsupervised DL model is proposed to directly convert CBCT images into sCT images with reduced artifacts and more accurate Hounsfield Unit values, as found in CT scans. Second, the model's generalizability is investigated and discussed, along with potential solutions to address the generalizability problem. Third, two DL models are designed to extract and combine information from pCT and CBCT, in order to inpaint axial and longitudinal truncations in CBCT images. Through these three studies, the predicted truncation-free sCT images hold the potential to enhance ART workflows, enabling more accurate dose calculations compared to DIR-generated sCT. The dissertation then shifts its focus to CBCT auto-segmentation. Initially, an unsupervised DL-based DIR model is proposed to predict the deformation vector field between pCT and CBCT, enabling pCT structure propagation as the segmentation on CBCT. To further improve segmentation accuracy, a DL-based direct segmentation model assisted by DIR is proposed, which outperforms state-of-the-art DIR-based segmentation results. The contributions of this thesis are expected to enhance the accuracy and efficiency of sCT generation and CBCT auto-segmentation in the CBCT-based ART workflow.