90 research outputs found
Progressive Seed Generation Auto-encoder for Unsupervised Point Cloud Learning
No full textWith the development of 3D scanning technologies, 3D vision tasks have become a popular research area. Owing to the large amount of data acquired by sensors, unsupervised learning is essential for understanding and utilizing point clouds without an expensive annotation process. In this paper, we propose a novel framework and an effective auto-encoder architecture named “PSG-Net” for reconstruction-based learning of point clouds. Unlike existing studies that used fixed or random 2D points, our framework generates input-dependent point-wise features for the latent point set. PSG-Net uses the encoded input to produce point-wise features through the seed generation module and extracts richer features in multiple stages with gradually increasing resolution by applying the seed feature propagation module progressively. We prove the effectiveness of PSG-Net experimentally; PSG-Net shows state-of-the-art performances in point cloud reconstruction and unsupervised classification, and achieves comparable performance to counterpart methods in supervised completion
Angstrom-accuracy multilayer thickness determination using optical metrology and machine learning
No full textThe era of big data and cloud computing services has driven the demand for higher capacity and more compact semiconductor devices. As a result, semiconductor devices are moving from 2-D to 3-D. Most notably, three-dimensional (3D) NAND flash memory is the most successful 3D semiconductor device today. 3D NAND overcomes the spatial limitation of conventional planar NAND by stacking memory cells vertically. Since hundreds of vertically stacked semiconductor materials become the channel length in the final product, accurate thickness characterization is critical. In this paper, we propose a non-destructive multilayer thickness characterization method using optical measurements and machine learning. For a silicon oxide/nitride multilayer stack of >200 layers, we could predict the thickness of each layer with an average root-mean-square error (RMSE) of 1.6 Å. In addition, we could successfully classify normal and outlier devices using simulated data. We expect this method to be highly suitable for semiconductor fabrication processes
Machine learning-based automated classification of headache disorders using patient-reported questionnaires
No full text© Te Author(s) 2020. Classification of headache disorders is dependent on a subjective self-report from patients and its interpretation by physicians. We aimed to apply objective data-driven machine learning approaches
to analyze patient-reported symptoms and test the feasibility of the automated classification of
headache disorders. The self-report data of 2162 patients were analyzed. Headache disorders were
merged into five major entities. The patients were divided into training (n = 1286) and test (n = 876)
cohorts. We trained a stacked classifier model with four layers of XGBoost classifiers. The first layer
classified between migraine and others, the second layer classified between tension-type headache
(TTH) and others, and the third layer classified between trigeminal autonomic cephalalgia (TAC) and
others, and the fourth layer classified between epicranial and thunderclap headaches. Each layer
selected different features from the self-reports by using least absolute shrinkage and selection
operator. In the test cohort, our stacked classifier obtained accuracy of 81%, sensitivity of 88%,
69%, 65%, 53%, and 51%, and specificity of 95%, 55%, 46%, 48%, and 51% for migraine, TTH, TAC,
epicranial headache, and thunderclap headaches, respectively. We showed that a machine-learning
based approach is applicable in analyzing patient-reported questionnaires. Our result could serve as a
baseline for future studies in headache research.11Nsciescopu
Plasma membrane localization of MLC1 regulates cellular morphology and motility
No full textBackground: Megalencephalic leukoencephalopathy with subcortical cysts (MLC) is a rare form of infantile-onset leukodystrophy. The disorder is caused primarily by mutations of MLC1 that leads to a series of phenotypic outcomes including vacuolation of myelin and astrocytes, subcortical cysts, brain edema, and macrocephaly. Recent studies have indicated that functional interactions among MLC1, GlialCAM, and ClC-2 channels play key roles in the regulation of neuronal, glial and vascular homeostasis. However, the physiological role of MLC1 in cellular homeostatic communication remains poorly understood. In the present study, we investigated the cellular function of MLC1 and its effects on cell-cell interactions. Methods: MLC1-dependent cellular morphology and motility were analyzed by using confocal and live cell imaging technique. Biochemical approaches such as immunoblotting, co-immunoprecipitation, and surface biotinylation were conducted to support data. Results: We found that the altered MLC1 expression and localization led to a great alteration in cellular morphology and motility through actin remodeling. MLC1 overexpression induced filopodia formation and suppressed motility. And, MLC1 proteins expressed in patient-derived MLC1 mutants resulted in trapping in the ER although no changes in morphology or motility were observed. Interestingly knockdown of Mlc1 induced Arp3-Cortactin interaction, lamellipodia formation, and increased the membrane ruffling of the astrocytes. These data indicate that subcellular localization of expressed MLC1 at the plasma membrane is critical for changes in actin dynamics through ARP2/3 complex. Thus, our results suggest that misallocation of pathogenic mutant MLC1 may disturbs the stable cell-cell communication and the homeostatic regulation of astrocytes in patients with MLC. © 2019 The Author(s).1
Nonparametric statistical methods for image segmentation and shape analysis
No full textThesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2005.This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.Page 131 blank.Includes bibliographical references (p. 125-130).Image segmentation, the process of decomposing an image into meaningful regions, is a fundamental problem in image processing and computer vision. Recently, image segmentation techniques based on active contour models with level set implementation have received considerable attention. The objective of this thesis is in the development of advanced active contour-based image segmentation methods that incorporate complex statistical information into the segmentation process, either about the image intensities or about the shapes of the objects to be segmented. To this end, we use nonparametric statistical methods for modeling both the intensity distributions and the shape distributions. Previous work on active contour-based segmentation considered the class of images in which each region can be distinguished from others by second order statistical features such as the mean or variance of image intensities of that region. This thesis addresses the problem of segmenting a more general class of images in which each region has a distinct arbitrary intensity distribution. To this end, we develop a nonparametric information-theoretic method for image segmentation. In particular, we cast the segmentation problem as the maximization of the mutual information between the region labels and the image pixel intensities. The resulting curve evolution equation is given in terms of nonparametric density estimates of intensity distributions, and the segmentation method can deal with a variety of intensity distributions in an unsupervised fashion. The second component of this thesis addresses the problem of estimating shape densities from training shapes and incorporating such shape prior densities into the image segmentation process.(cont.) To this end, we propose nonparametric density estimation methods in the space of curves and the space of signed distance functions. We then derive a corresponding curve evolution equation for shape-based image segmentation. Finally, we consider the case in which the shape density is estimated from training shapes that form multiple clusters. This case leads to the construction of complex, potentially multi-modal prior densities for shapes. As compared to existing methods, our shape priors can: (a) model more complex shape distributions; (b) deal with shape variability in a more principled way; and (c) represent more complex shapes.by Junmo Kim.Ph.D
U-Net-Based Segmentation for Electrical Lines and Its Application to Real-Time Maintenance Algorithm for Electricity Facilities
No full textInspections for maintenance of electricity facility are performed by patrols which are always in a potentially dangerous traffic accident. In this study, we address the inspection procedure with leveraging advanced deep learning algorithms on images for patrollers to focus on driving to reduce the danger, assuming that facility images are obtained by photographing automatically with a monocular camera during patrols. Toward the goal with restriction of our concerns on an electrical line and related devices, it is initially proposed a new image segmentation algorithm for electrical lines, U-Net-based CNN model, which can be used as a basic step of real-time maintenance algorithms for electricity facilities. It is then introduced a novel inference algorithm for the connectivity between an insulator and an electrical line, as one realization of real-time maintenance algorithms. Experiments demonstrate that the both proposed algorithms are effective and efficient enough to be feasible for real-time algorithms. The proposed methods expect not only to help dangerous labor done by patrollers but also to save lots of money and time by products. © 2021, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd
Disrupted stepwise functional brain organization in overweight individuals
No full text© 2022, The Author(s).Functional hierarchy establishes core axes of the brain, and overweight individuals show alterations in the networks anchored on these axes, particularly in those involved in sensory and cognitive control systems. However, quantitative assessments of hierarchical brain organization in overweight individuals are lacking. Capitalizing stepwise functional connectivity analysis, we assess altered functional connectivity in overweight individuals relative to healthy weight controls along the brain hierarchy. Seeding from the brain regions associated with obesity phenotypes, we conduct stepwise connectivity analysis at different step distances and compare functional degrees between the groups. We find strong functional connectivity in the somatomotor and prefrontal cortices in both groups, and both converge to transmodal systems, including frontoparietal and default-mode networks, as the number of steps increased. Conversely, compared with the healthy weight group, overweight individuals show a marked decrease in functional degree in somatosensory and attention networks across the steps, whereas visual and limbic networks show an increasing trend. Associating functional degree with eating behaviors, we observe negative associations between functional degrees in sensory networks and hunger and disinhibition-related behaviors. Our findings suggest that overweight individuals show disrupted functional network organization along the hierarchical axis of the brain and these results provide insights for behavioral associations.11Nsciescopu
Radiomics-guided deep neural networks stratify lung adenocarcinoma prognosis from CT scans
No full text© 2021, The Author(s).Deep learning (DL) is a breakthrough technology for medical imaging with high sample size requirements and interpretability issues. Using a pretrained DL model through a radiomics-guided approach, we propose a methodology for stratifying the prognosis of lung adenocarcinomas based on pretreatment CT. Our approach allows us to apply DL with smaller sample size requirements and enhanced interpretability. Baseline radiomics and DL models for the prognosis of lung adenocarcinomas were developed and tested using local (n = 617) cohort. The DL models were further tested in an external validation (n = 70) cohort. The local cohort was divided into training and test cohorts. A radiomics risk score (RRS) was developed using Cox-LASSO. Three pretrained DL networks derived from natural images were used to extract the DL features. The features were further guided using radiomics by retaining those DL features whose correlations with the radiomics features were high and Bonferroni-corrected p-values were low. The retained DL features were subject to a Cox-LASSO when constructing DL risk scores (DRS). The risk groups stratified by the RRS and DRS showed a significant difference in training, testing, and validation cohorts. The DL features were interpreted using existing radiomics features, and the texture features explained the DL features well.11Nsciescopu
포인트 클라우드 복원과 지식 증류를 통한 3차원 데이터 및 모델 압축
No full text학위논문(박사) - 한국과학기술원 : 전기및전자공학부, 2023.2,[iii, 44 p. :]Following considerable development in 3D scanning technologies, many deep learning studies have recently been proposed with various approaches for 3D vision tasks. Also, in order to make practical use of these methods, the importance of a lightweight model with an efficient system is increasing. Therefore, we focus on data compression and model compression in 3D vision. In this thesis, we firstly proposed a novel framework and an effective auto-encoder architecture for data compression. Unlike existing studies that used fixed or random 2D points, our framework facilitates point cloud reconstruction by generating input-dependent point-wise features for the latent point set. Our method shows state-of-the-art performances in point cloud reconstruction and unsupervised classification, and achieves comparable performance to counterpart methods in supervised completion. For model compression, we revisit the basic concept of knowledge distillation and compare three losses derived from measures that calculate the similarity between the teacher prediction and the student prediction: Kullback-Leibler divergence (KLD), mean squared error (MSE), and cosine similarity (CS) losses. Unlike previous studies concerned with KLD and MSE losses that transfer the teacher logit values, we explored the possibility of the CS loss transferring the direction of the teacher logit. The CS loss achieved performance comparable to state-of-the-art with superior efficiency in terms of training time and the number of parameters. We finally applied knowledge distillation using the CS loss to 3D models to perform model compression.한국과학기술원 :전기및전자공학부
A Cooperative Protocol for Vehicle Merging Using Bi-dimensional Artificial Potential Fields
No full textIn recent years, platooning solutions like cooperative adaptive cruise control (CACC) have been deeply studied. It is common in such platooning literature to assume that the vehicles drive on the same lane (longitudinal platooning). At the same time, lateral control during merging maneuvers is commonly addressed as a path planning problem, in which the ego vehicle changes the lane during merging without necessarily cooperating with its neighboring vehicles (i.e. without considering gap closing). The primary objective of this article is to develop a control strategy which involves both longitudinal and lateral vehicle dynamics, where the vehicles merge and form a platoon in a cooperative way without a priori path planning. Appropriately designed bi-dimensional artificial potential fields are used to achieve this goal and the proposed protocol is verified through simulations with CarSim.Accepted Author ManuscriptTeam Bart De Schutte
- …
