Raw Scan Research Articles

Robust Lidar-Based Localization Scheme for Unmanned Ground Vehicle via Multisensor Fusion.

This article proposes a robust and precise localization scheme for unmanned ground vehicle (UGV) in global positioning system (GPS)-denied and GPS-challenged environments via multisensor fusion approach. The localization scheme is proposed to be under an available point-cloud map. First, initialization in localization module is designed to calculate the initial position of UGV in map using the Gaussian projection approach and obtain the frame transformation between the 3-D lidar and the inertial measurement unit (IMU). Second, the best alignment between each scan frame and the available submap is obtained, and the pose of vehicle relative to the origin of map is calculated. Third, the precise pose of UGV is well predicted by integrating the data from 3-D lidar and IMU. Fourth, in order to visually verify the proposed localization scheme, the motion of vehicle is visualized by designing the visualization module. Note that the preprocessing module aims to process the raw scan data. The availability of our proposed localization scheme is verified by conducting experiments in our campus.

Skeleton extraction from point clouds of trees with complex branches via graph contraction

Extracting a skeleton from a 3D tree-shaped point cloud with complex branches is a challenging issue due to the diversity of branches and their natural topological complexity. In this paper, we first introduce a novel contraction method called “graph contraction” to contract 3D tree-shaped point clouds. The computation of the graph contraction is formulated as the minimization of an energy function that consists of a contraction term that minimizes the sum of the graph geodesic distances of the k-nearest geodesic neighbors and a topology-preserving term that prevents point clouds from shrinking in the local principal direction. After graph contraction, we downsample the initial skeleton and obtain skeleton nodes. Finally, we introduce an algorithm that can be used to extract a topologically correct skeleton by connecting skeleton nodes. The proposed method has been validated on various 3D tree-shaped point clouds, including reconstructed tree-shaped point clouds, artificially generated tree-shaped point clouds and raw scan data. We use the Chamfer distance to measure the correctness of the skeleton. The experimental results show that the Chamfer distance obtained using our method is smaller than that obtained using current state-of-the-art methods for skeleton extraction of 3D tree-shaped point clouds with complex branches.

Feasibility study of a smartphone pupillometer and evaluation of its accuracy.

Measurement of pupillary characteristics, such as pupillary unrest in ambient light, and reflex dilation have been shown to be useful in a variety of clinical situations. Dedicated pupillometers typically capture images in the near-infrared to allow imaging in both light and darkness. However, because a subset of pupillary measurements can be acquired with levels of visible light suitable for conventional cameras, it is theoretically possible to capture data using general purpose cameras and computing devices such as those found on smartphones. Here we describe the development of a smartphone-based pupillometer and compare its performance with a commercial pupillometer. Smartphone pupillometry software was developed and then compared with a commercial pupillometer by performing simultaneous scans in both eyes, using the smartphone pupillometer and a commercial pupillometer. The raw scans were compared, as well as a selected pupillary index: pupillary unrest in ambient light. In 77% of the scans the software was able to successfully identify the pupil and iris. The raw data as well as calculated values of pupillary unrest in ambient light were in clinically acceptable levels of agreement; Bland-Altman analysis of raw pupil measurements yielded a 95% confidence interval of 0.26mm. In certain situations a smartphone pupillometer may be an appropriate alternative to a commercial pupillometer.

A bivariate behavioural genetics model of sleep and cortisol in middle childhood

This report evaluated sex-specific differences in the association between brain aging and body mass index (BMI) in young adults using the publicly available data from the Human Connectome Project (HCP).Participants of HCP with available structural imaging and BMI data were included [n = 1112; mean age = 28.80 (SD = 3.70); mean BMI = 26.53 (SD = 5.20); males n = 507, females n = 605]. Predicted brain age was generated using raw T1-weighted MRI scan and a Gaussian Processes regression model. The difference (Δ aging) between brain age predicted by structural imaging and chronological age was computed. A linear regression model was used with Δ aging as the dependent variable, and sex, BMI, and BMI-by-sex interaction as independent variables of interest, and race, ethnicity, income, and education as covariates.There was a significant BMI-by-sex interaction for Δ aging (p = 0.041). Higher BMI was associated with greater brain aging in both sexes. However, this association was substantially stronger in males (β = 0.215; SE = 0.050; p < 0.0001) than in females (β = 0.122; SE = 0.035; p = 0.0005).We found evidence suggesting that higher BMI is associated with greater brain aging in adults. Furthermore, the association between higher BMI and greater brain aging was stronger in males than in females. Future studies are needed to explore the mechanistic pathways that link higher BMI to greater brain aging and whether weight-loss interventions, such as exercise, can reverse higher BMI-associated greater brain aging.

Synthetic Aperture Sonar (SAS) without Navigation: Scan Registration as Basis for Near Field Synthetic Imaging in 2D.

Sonars are essential for underwater sensing as they can operate over extended ranges and in poor visibility conditions. The use of a synthetic aperture is a popular approach to increase the resolution of sonars, i.e., the sonar with its N transducers is positioned at k places to generate a virtual sensor with transducers. The state of the art for synthetic aperture sonar (SAS) is strongly coupled to constraints, especially with respect to the trajectory of the placements and the need for good navigation data. In this article, we introduce an approach to SAS using registration of scans from single arrays, i.e., at individual poses of arbitrary trajectories, hence avoiding the need for navigation data of conventional SAS systems. The approach is introduced here for the near field using the coherent phase information of sonar scans. A Delay and Sum (D&S) beamformer (BF) is used, which directly operates on pixel/voxel form on a Cartesian grid supporting the registration. It is shown that this pixel/voxel-based registration and the coherent processing of several scans forming a synthetic aperture yields substantial improvements of the image resolution. The experimental evaluation is done with an advanced simulation tool generating realistic 2D sonar array data, i.e., with simulations of a linear 1D antenna reconstructing 2D images. For the image registration of the raw sonar scans, a robust implementation of a spectral method is presented. Furthermore, analyses with respect to the trajectories of the sensor locations are provided to remedy possible grating lobes due to the gaping positions of the transmitter devices.

STRUCTURE-FROM-MOTION PROCESSING OF AERIAL ARCHIVE PHOTOGRAPHS: SENSITIVITY ANALYSES PAVE THE WAY FOR QUANTIFYING GEOMORPHOLOGICAL CHANGES SINCE 1978 IN LA RÉUNION ISLAND

Abstract. Landscapes have been photographed dozens of times at scales ca. 1/25,000 and better since World War II. Scans are distributed freely online (e.g. remonterletemps.ign.fr). In parallel, Structure-from-Motion (SFM) software made photogrammetric processing easy to non-specialists. Yet puzzling questions crop up to use both: (i) Can raw scans be used as is? (ii) Can Ground Control Points (GCP) and checkpoints be safely collected from a web portal? (iii) How many parameters are sufficient for camera interior orientation? (iv) Are single flight camera networks sufficient to constrain camera models compared to multiple flights? (v) Are photogrammetric Digital Surface Models (DSM) fit for quantifying landslide activity? Processing of scanned black-and-white 1/27,000 photographs from IGN flown in May 1978 over Cirque de Salazie in La Réunion Island answer these questions. We find that raw scanned photographs need translation, rotation and cropping to match the camera reference frame. GCP and Check point coordinates collected on geoportail.gouv.fr with assumed accuracy of 10 m, achieved ca. 7 m accurate SFM registration. The optimal camera model uses only 4 parameters: f, cx, cy and K1. Compared to a 2015 lidar Digital Terrain Model (DTM), the 0.66 m/pixel DSM of 1978 has a median deviation of −1.39 m ± 3.34 m (Median Absolute Deviation) which is comparable to GCP quality. Elevation difference more importantly reveals, for the first time, the 37 years and 13 cyclones cumulated landslides pattern on Cirque de Salazie. Photographic archives hold decades-long 3D history. SFM is a game changer for landslide risk mitigation planning.

Lipid profile dataset of optogenetics induced optic nerve regeneration.

The optic nerve transfers visual information from the retina to the brain through the axons of retinal ganglion cells (RGCs). In adult mammals, optic nerve injuries and progressive degenerative diseases lead to the irreversible loss of RGCs, resulting in vision loss and blindness. Optogenetic models have proved useful in manipulating the growth of RGCs through expression and stimulation of channelrhodopsins (Chr2) in RGCs using the RGC-specific thy-1 promoter. Using transgenic Chr2 mouse (Thy1-ChR2-EYFP) as a model of regeneration, we profile the lipid changes which occur after traumatic optic nerve crush, light stimulation and forced RGC axonal growth. Thy1-ChR2-EYFP and control (C57BL/6) mice were divided in four groups each – 1) no crush and no stimulation, 2) no crush with stimulation, 3) crush and without stimulation, and 4) crush with stimulation. After euthanasia, the optic nerves were collected for lipidomic analysis. The Bligh and Dyer method was used for lipid extraction, followed by mass spectrometry lipid profiling with a Q-Exactive Orbitrap Liquid Chromatography-Mass Spectrometer (LC MS-MS). The raw scans were analysed with LipidSearch 4.1.3 and the statistical analysis was conducted through Metaboanalyst 4.0. This data is available at Metabolomics Workbench, study ID ST001381: [https://www.metabolomicsworkbench.org/data/DRCCMetadata.php?Mode=Study&StudyID=ST001381&StudyType=MS&ResultType=5].

Lumbar Vertebrae Synthetic Segmentation in Computed Tomography Images Using Hybrid Deep Generative Adversarial Networks.

The lumbar vertebrae segmentation in Computed tomography (CT) is challenging due to the scarcity of the labeled training data that we define as paired training data for the deep learning technique. Much of the available data is limited to the raw CT scans, unlabeled by radiologists. To handle the scarcity of labeled data, we utilized a hybrid training system by combining paired and unpaired training data and construct a hybrid deep segmentation generative adversarial network (Hybrid-SegGAN). We develop a total automatic approach for lumbar vertebrae segmentation in CT images using Hybrid-SegGAN for synthetic segmentation. Our network receives paired and unpaired data, discriminates between the two sets of data, and processes each through separate phases. We used CT images from 120 patients to demonstrate the performance of the proposed method and extensively evaluate the segmentation results against their ground truth by using 12 performance measures. The result analysis of the proposed method suggests its feasibility to improve the capabilities of deep learning segmentation without demanding the time-consuming annotation procedure for labeled and paired data.

인간공학적 제품 설계를 위한 3차원 탬플릿 인체 모델의 활용

Objective: This study aims to provide a comprehensive review of the literature on template model registration technologies and their applications in ergonomic product design. Background: The original 3D scan data of the human body has limitations in terms of its applicability due to lacking quality of mesh (e.g., missed scan areas and noise) and inconsistent postures between participants. Template model registration technologies are used to complement the lacking quality of raw 3D scan data in anthropometric research and ergonomic product design. Method: A literature review was conducted on 66 papers, which were selected by screening titles and abstracts of papers searched by keywords including template model, template matching, and template registration within ergonomics and computer graphics domains in Scopus. Results: A human template model generally consists of five elements: (1) point-cloud and mesh, (2) reference points, (3) skeletal structure of the body, (4) body segment, and (5) skin deformation characteristics. The template model is matched to a target image with a particular size, shape, and posture using a template model registration technology. Conclusion: The template model registration technology has a great potential in its applicability to ergonomic design and evaluation of products for a comfortable fit to the human body. Application: The template model registration technology can be effectively applied to the establishment of parameterized 3D human body database, automatic measurement of body sizes, analysis of the variation of body shapes, predictive analysis of the body posture, and customized design of a product.

Mind your privacy: Privacy leakage through BCI applications using machine learning methods

With the digitization of almost every aspect of our lives, privacy leakage in cyber space has become a pressing concern. Brain–Computer Interface (BCI) systems have become more popular in recent years and are now being used for a variety of applications. BCI data represents an individual’s brain activity at a given time. Like many other kinds of data, BCI data can be utilized for malicious purposes. Electroencephalography (EEG) is one of the most popular brain activity acquisition methods of BCI applications. More specifically, BCI games, represent one of the main EEG applications. However, a malicious BCI application (e.g. game) could allow an attacker to take advantage of an unsuspecting user happily enjoying a game and record the user’s brain activity; by analyzing this data, the attacker can infer private information and characteristics regarding the user, without his/her consent or awareness. This study is the first to demonstrate the ability to predict and infer meaningful personality traits and cognitive abilities by analyzing resting-state EEG (rsEEG) recordings of an individual’s brain activity using a variety of machine learning methods. A comprehensive set of raw rsEEG scans, along with the dissociation level and executive function (EF) performance measures, for the 162 subjects were used in our evaluation. The best results we achieved were an accuracy of 73% for dissociation classification and less than 16% mean absolute error in predicting performance for all examined EFs. These encouraging results are better than those presented in prior research, both in terms of accuracy and data-validity and dataset size.

Quantification of Retinal Nerve Fibre Layer Thickness on Optical Coherence Tomography with a Deep Learning Segmentation-Free Approach

This study describes a segmentation-free deep learning (DL) algorithm for measuring retinal nerve fibre layer (RNFL) thickness on spectral-domain optical coherence tomography (SDOCT). The study included 25,285 B-scans from 1,338 eyes of 706 subjects. Training was done to predict RNFL thickness from raw unsegmented scans using conventional RNFL thickness measurements from good quality images as targets, forcing the DL algorithm to learn its own representation of RNFL. The algorithm was tested in three different sets: (1) images without segmentation errors or artefacts, (2) low-quality images with segmentation errors, and (3) images with other artefacts. In test set 1, segmentation-free RNFL predictions were highly correlated with conventional RNFL thickness (r = 0.983, P < 0.001). In test set 2, segmentation-free predictions had higher correlation with the best available estimate (tests with good quality taken in the same date) compared to those from the conventional algorithm (r = 0.972 vs. r = 0.829, respectively; P < 0.001). Segmentation-free predictions were also better in test set 3 (r = 0.940 vs. r = 0.640, P < 0.001). In conclusion, a novel segmentation-free algorithm to extract RNFL thickness performed similarly to the conventional method in good quality images and better in images with errors or other artefacts.

Multi-View Features Joint Learning with Label and Local Distribution Consistency for Point Cloud Classification

In outdoor Light Detection and Ranging (lidar)point cloud classification, finding the discriminative features for point cloud perception and scene understanding represents one of the great challenges. The features derived from defect-laden (i.e., noise, outliers, occlusions and irregularities) and raw outdoor LiDAR scans usually contain redundant and irrelevant information which adversely affects the accuracy of point semantic labeling. Moreover, point cloud features of different views have a capability to express different attributes of the same point. The simplest way of concatenating these features of different views cannot guarantee the applicability and effectiveness of the fused features. To solve these problems and achieve outdoor point cloud classification with fewer training samples, we propose a novel multi-view features and classifiers’ joint learning framework. The proposed framework uses label consistency and local distribution consistency of multi-space constraints for multi-view point cloud features extraction and classification. In the framework, the manifold learning is used to carry out subspace joint learning of multi-view features by introducing three kinds of constraints, i.e., local distribution consistency of feature space and position space, label consistency among multi-view predicted labels and ground truth, and label consistency among multi-view predicted labels. The proposed model can be well trained by fewer training points, and an iterative algorithm is used to solve the joint optimization of multi-view feature projection matrices and linear classifiers. Subsequently, the multi-view features are fused and used for point cloud classification effectively. We evaluate the proposed method on five different point cloud scenes and experimental results demonstrate that the classification performance of the proposed method is at par or outperforms the compared algorithms.

Lung Nodule Texture Detection and Classification Using 3D CNN

Following artificial intelligence implementation in computer vision fields, especially deep learning, many Computer-Aided Diagnosis (CAD) tools are proposed to help to detect lung cancer by the scoring system or by identifying the characteristics of nodules. However, lung cancer is a clinical diagnosis which does not provide detailed information needed by radiologists and clinician to prevent unnecessary invasive diagnostic procedures compared to lung nodule texture detection and classification. Hence, to answer this problem, this research explores the steps needed to implement 3D CNN on raw thorax CT scan datasets and usage of RetinaNet 3D + Inception 3D with transfer learning. The 3D CNN CAD tools can improve the speed, performance, and ability to detect lung nodule texture instead of malignancy status done by previous studies. This study implements 3D CNN on Moscow private datasets acquired from NVIDIA Asia Pacific. The proposed method of data conversion can minimize information loss from raw data to 3D CNN input data. After 100 epoch training, the researchers conclude that the best-proposed model (3D CNN with transfer learning of pretrained LIDC public datasets weight) provides 22.86% of mean average precision (mAP) detection capability and 70.36% of Area Under the Curve (AUC) in Moscow private dataset lung texture detection tasks. It outperforms non-transfer learning 3D CNN model (trained from scratch) and 3D CNN with transfer learning of pre-trained ImageNet weight.

NRLI-UAV: Non-rigid registration of sequential raw laser scans and images for low-cost UAV LiDAR point cloud quality improvement

Accurate registration of light detection and ranging (LiDAR) point clouds and images is a prerequisite for integrating the spectral and geometrical information collected by low-cost unmanned aerial vehicle (UAV) systems. Most registration approaches take the directly georeferenced LiDAR point cloud as a rigid body, based on the assumption that the high-precision positioning and orientation system (POS) in the LiDAR system provides sufficient precision, and that the POS errors are negligible. However, due to the large errors of the low-precision POSs commonly used in the low-cost UAV LiDAR systems (ULSs), dramatic deformation may exist in the directly georeferenced ULS point cloud, resulting in non-rigid transformation between the images and the deformed ULS point cloud. As a result, registration may fail when using a rigid transformation between the images and the directly georeferenced LiDAR point clouds. To address this problem, we proposed NRLI-UAV, which is a non-rigid registration method for registration of sequential raw laser scans and images collected by low-cost UAV systems. NRLI-UAV is a two-step registration method that exploits trajectory correction and discrepancy minimization between the depths derived from structure from motion (SfM) and the raw laser scans to achieve LiDAR point cloud quality improvement. Firstly, the coarse registration procedure utilizes global navigation satellite system (GNSS) and inertial measurement unit (IMU)-aided SfM to obtain accurate image orientation and corrects the errors of the low-precision POS. Secondly, the fine registration procedure transforms the original 2D-3D registration to 3D-3D registration. This is performed by setting the oriented images as the reference, and iteratively minimizing the discrepancy between the depth maps derived from SfM and the raw laser scans, resulting in accurate registration between the images and the LiDAR point clouds. In addition, an improved LiDAR point cloud is generated in the mapping frame. Experiments were conducted with data collected by a low-cost UAV system in three challenging scenes to evaluate NRLI-UAV. The final registration errors of the images and the LiDAR point cloud are less than one pixel in image space and less than 0.13 m in object space. The LiDAR point cloud quality was also evaluated by plane fitting, and the results show that the LiDAR point cloud quality is improved by 8.8 times from 0.45 m (root-mean-square error [RMSE] of plane fitting) to 0.05 m (RMSE of plane fitting) using NRLI-UAV, demonstrating a high level of automation, robustness, and accuracy.

Structure-consistent customized virtual mannequin reconstruction from 3D scans based on optimization

The 3D virtual mannequin has been widely used in apparel industry, and its importance is also increasing. This work develops a new 3D virtual mannequin reconstruction system based on optimization. All the mannequins reconstructed by the proposed approach share the identical topology, that is, there is a point-to-point correspondence among the mannequins, which will significantly facilitate much subsequent processing in fashion design, made-to-measure, and virtual try-on. The inputs to the proposed system contain a template human body, a raw scan (represented in mesh), and a very sparse corresponding landmarks set. The proposed approach substantially utilizes the optimization technology to drive the template to deform into a real scan. There is no special requirement on the raw meshes. The raw meshes may have a different number of vertices and triangles or may even be incomplete. The proposed method only needs 21 landmarks as hard-constraints to reconstruct a mannequin with tens of thousands of vertices. These landmarks can be extracted automatically for standard mannequin reconstruction. Besides the standard mannequin, the proposed system can also be used to reconstruct display mannequins, that is, mannequins with various poses. The experiments visualize the optimization procedure and verify that the optimization is efficient and effective. Quantitative analysis also proves that the reconstruction error satisfies the requirements of fashion design and tailoring.

Custom Optical Coherence Tomography Parameters for Distinguishing Papilledema from Pseudopapilledema.

Causes of papilledema can be life-threatening; however, distinguishing papilledema from pseudopapilledema is often challenging. The conventional optical coherence tomography (OCT) scan for assessing the optic nerve often fails to detect mild papilledema. Our study suggests that parameters derived from volumetric OCT scans can provide additional useful information for detecting papilledema. Optical coherence tomography analysis of the optic nerve commonly measures retinal nerve fiber layer thickness (RNFLT) along a 1.73-mm-radius scan path. This conventional scan, however, often fails to detect mild papilledema. The purpose of this study was to evaluate additional OCT-derived measures of the optic nerve head (ONH) and peripapillary retina for differentiating papilledema (all grades and mild) from pseudopapilledema. Cirrus OCT ONH volume scans were acquired from 21 papilledema (15 mild papilledema), 27 pseudopapilledema, and 42 control subjects. Raw scan data were exported, and total retinal thickness within Bruch's membrane opening (BMO) plus RNFLT and total retinal thickness at the following eccentricities were calculated using custom algorithms: BMO to 250, 250 to 500, 500 to 1000, and 1000 to 1500 μm. Minimum rim width was calculated, and BMO height was measured from a 4-mm Bruch's membrane reference plane centered on the BMO. Retinal nerve fiber layer thickness from BMO to 250 μm, minimum rim width, and BMO height had significantly greater areas under the receiver operating characteristic curve than did conventional RNFLT for differentiating mild papilledema from pseudopapilledema (P < .0001) and greater sensitivities at 95% specificity. Using cutoff values at 95% specificity, custom parameters detected 10 mild papilledema patients, and conventional RNFLT detected only 1. Bruch's membrane opening heights above the reference plane were observed in papilledema only, although many papilledema cases had a neutral or negative BMO height. Using OCT volumetric data, additional parameters describing peripapillary tissue thickness, neuroretinal rim thickness, and ONH position can be calculated and provide valuable measures for differentiating mild papilledema from pseudopapilledema.

APPLICATIONS OF HIPPOCAMPUS ANATOMICAL DEFINITIONS AND MEASUREMENT METHODS IN THE USE AS A BIOMARKER

Hippocampus measures are one of the staple biomarkers used in following the progression of amnestic mild cognitive impairment and Alzheimer's disease. However, the anatomical definition and methods vary considerably, and may have a direct impact on ability to detect significant changes. For studies where it was available, measures were used from the ADNI site. This included SNT, FreeSurfer 5.3 and NeuroQuant. In addition, raw scans downloaded from ADNI were processed using FSL, NeuroAnalytica 3.10 and FreeSurfer 6.0. In addition, manual segmentations available from the Harmonized Hippocampus Protocol were included in the comparisons. In the comparisons only data that was common to all were used. Hippocampus volumes for the samples ranged widely depending on the anatomical definition used, with the largest being ∼80% larger than the smallest. Regression between the methods showed a non-zero intercept reflecting the different anatomical definitions, but correlation was generally quite good. Group analyses between AD, MCI and controls using each of the methods show relatively similar differences between groups, regardless of the anatomical definition or methods used. Power analysis to assess the use of 12-month atrophy measures to monitor progression yielded widely different results, with N80 (the number of subjects per arm of a clinical trial to detect a 25% reduction in atrophy with 80% power and 95% confidence) ranging from 52 to over 300. Correlation of hippocampus volume with memory scores ranged from 0.18 to 0.28 (R2) depending on both the methods used and the anatomical definition. For simple group comparisons, most hippocampus anatomical definitions and measurement methods provide adequate results. However, for use in evaluation of cognitive correlates and monitoring atrophy, the selection of both anatomical definition and measurement method are important in the success of the study.

Hierarchical tunnel modeling from 3D raw LiDAR point cloud

Precise modeling of tunnel structures can be used to evaluate the performance and state of safety as an important construction analysis object. However, the problem of modeling on a large-scale tunnel is challenging, due to the poor-quality of the input data that is contaminated with severe missing areas, noise, outliers and the accessories. This work introduces the concept of hierarchical modeling to automatically derive a measured and accurate representation of a 3D shield tunnel. Our key observation is that the shield tunnel is composed of a certain number of segment pieces arranged in some designed patterns. We hereby apply hierarchical segmentation and element extraction, based on designed patterns, to intensity images without expensive computation. According to the corresponding relationship between the point cloud and the image, the segment pieces within the point cloud are extracted. Therefore, we construct the corresponding 3D models of the segment pieces to obtain the precise structure model of the tunnel. As a result, segment pieces are extracted robustly, even in the presence of noise and large occlusion. Quantitative and qualitative comparisons of the proposed algorithm are presented with state-of-the-art methods. We evaluate the modeling method on a variety of raw LiDAR scans, in terms of its robustness and accuracy. Moreover, our method has been successfully applied in a number of practical projects to discover the multi-level structures of the shield tunnels, which are used to evaluate the safety state of tunnels.

Comparison of advanced discretization techniques for image-based modelling of heterogeneous porous rocks

This contribution presents an assessment of computational techniques enabling automated simulations of complex porous rocks microstructures based on 3D imaging techniques. A subset of a CT-scanned sandstone sample is used to compare the results obtained by two advanced discretization frameworks. Raw scan results are processed by a level-set-based segmentation technique to produce smooth geometries prone to finite element discretizations. A recently developed technique is outlined for conforming mesh generation for complex porous geometries described implicitly by functions. This allows generating high-quality tetrahedral meshes with selective refinement. Next to this, a technique that uses a kinematic enrichment by incompatible modes to represent the heterogeneous geometry is explained. Both techniques use the same implicit geometry as main input for the simulations. Mechanical simulations are conducted on a subset of a scanned sample of a sandstone under triaxial loading conditions for isotropic compressive loading and for loading conditions involving a stress deviator. The results are compared and discussed based on local stress distributions and on a Mohr–Coulomb criterion with tensile cut-off. The results show that both discretization strategies yield complementary tools and allow envisioning automated simulations based on raw CT scan data for porous rocks exhibiting complex pore space morphologies.

Fast and fully-automated detection and segmentation of pulmonary nodules in thoracic CT scans using deep convolutional neural networks

Deep learning techniques have been extensively used in computerized pulmonary nodule analysis in recent years. Many reported studies still utilized hybrid methods for diagnosis, in which convolutional neural networks (CNNs) are used only as one part of the pipeline, and the whole system still needs either traditional image processing modules or human intervention to obtain final results. In this paper, we introduced a fast and fully-automated end-to-end system that can efficiently segment precise lung nodule contours from raw thoracic CT scans. Our proposed system has four major modules: candidate nodule detection with Faster regional-CNN (R-CNN), candidate merging, false positive (FP) reduction with CNN, and nodule segmentation with customized fully convolutional neural network (FCN). The entire system has no human interaction or database specific design. The average runtime is about 16 s per scan on a standard workstation. The nodule detection accuracy is 91.4% and 94.6% with an average of 1 and 4 false positives (FPs) per scan. The average dice coefficient of nodule segmentation compared to the groundtruth is 0.793.