Showing 13–24 of 183 results

  • Borehole Imaging by Applying 3D Visual SLAM to Borehole Images Acquired in Forward Vision Camera System

    Abstract: In the drilling survey, there is a technique called borehole imaging, which examines the
    internal ground structure and the presence or absence of faults based on the unfolded panorama created by processing the borehole images taken by a camera. When we investigate the underground to prevent disasters such as landslides, it is desirable to use a portable and simple system including a monocular camera because we set up the system on the mountain slope. The problem with the monocular camera system is that the camera moves off the hole center and rotates around the hole axis. In order to generate the accurate unfolded images, it is necessary to obtain information about the accurate center and orientation of the camera moving inside the hole. For this purpose, we have developed a system to simultaneously estimate the camera pose and the hole center by applying SLAM technology to forward-looking images acquired in a forward vision camera system.

    Keywords: borehole imaging, panorama, visual SLAM, forward-looking images, monocular camera, optical flow, disaster prevention engineering, drilling survey.

  • EEG Signal Power Prediction Using DEAP Dataset

    Abstract: When we listen to music, our emotions can change due to changes in our brain response. Because of this, a large number of research projects for classifying the emotional response using brain signals when listening to music through machine learning techniques emerged. On the contrary, to our knowledge, there is no previous research attempting to estimate the dynamic changes in the electroencephalogram (EEG) response under music stimuli through machine learning techniques. Therefore, in this manuscript, we proposed an approach to predict and anticipate changes in the EEG signal under music stimuli. Using the DEAP dataset, we split the EEG response to music stimuli into one-second length frames. After that, we compared the changes in the power of the brain signal of consecutive frames through two one-tailed Wilcoxon rank-sum tests. This test allowed us to label the changes in the second frame as “lower”, “similar” or “higher” signal compared to the first frame. Then, we attempted to predict these changes using a Support-Vector Machine (SVM) classifier with stratified 5-fold validation with different input combinations (only music, only brain signal, or a combination of both). Due to the use of multi-label classification with imbalanced data, we measured the results through F1-Scores. Over chance level predictions of the changes of signal power were obtained when using the previous second brain signal for the different channels and bands, especially in the frontal F3 and F4 channels.

    Keywords: Electroencephalography (EEG), Music perception, Neural correlation, Machine learning, Signal classification.

  • Mixed-Precision Acceleration of Ultrasound Imaging Software

  • Evaluation of 2D and 3D Deep Learning Approaches for Automatic Segmentation of the Retinal External Limiting Membrane in Spectral Domain Optical Coherence Tomography Images

    Abstract: In this work, we compared the performance of 2D and 3D versions of three state-of-the-art deep neural networks on segmenting the retinal external limiting membrane (ELM) using a publicly available image dataset of spectral-domain optical coherence tomography (OCT) scans. Based on our results, 3D networks generally outperformed the 2D networks in Dice coefficient, mean surface distance and false positive rate but lagged behind in Hausdorff distance. 3D networks also produce smoother surfaces based on mean surface curvedness.

    Keywords: 2D vs. 3D Image Segmentation, Machine Learning, Annotation, External Limiting Membrane, OCT.

  • A sparse-NTF-based feature space construction method for automatic sleep stage identification

    Abstract: Healthy sleep is essential for the normal functioning of the human body and for the maintenance of mental vitality. One’s overnight sleep is usually evaluated by several sleep stages with sleep cycles. Automatic sleep stage identification methods are effective tools for sleep staging where feature extraction is an important procedure affecting the identification performance. In this study, an automatic sleep staging method based on multi-channel EEG signals and its optimized tensor feature space is developed. Several characteristic features are calculated from the two EEG recording channels and constructed as the original tensor feature space. A non-negative tensor factorization method based on sparse improvement is developed to optimize the tensor feature space for sleep staging. A classification model is constructed based on BP neural network and the parameters are estimated by PSO algorithm. Totally 10 overnight sleep recordings were tested. The averaged classification accuracy is about 84%. The developed method can be an assistant computerized tool for sleep staging.

    Keywords: Sleep staging; EEG; Non-negative tensor factorization; BP neural network; PSO

  • Deploying Electronic Health Records In Sports

    Abstract: EHR has the advantage of providing and updating information instantaneously at any time and place, giving it a potential advantage for use in sports events and not just in hospital settings.  With this in mind, we address the question: do frontline sports health providers, coaches, and athletes prefer traditional paper charts, or do they prefer a paperless version of the incident report, such as EHR? The research was conducted over three days at an organized outdoor summer sporting event, where medical providers used the EHR (Diya Sports Suite) platform to document the treatment of athletes plus spectators.  Data was collected via post-event survey questionnaires from coaches, providers, and scribes to solicit their feedback on the EHR.  For the medical providers, 86% of them feel that using EHR makes the treatment experience less stressful than traditional paperwork.  Overall, 86% of the providers, 62% of the coaches, and 100% of the scribes prefer a paperless medical record system over the traditional paper recording system.  However, this study was limited by the low number (12 documentation) of patient incidents seen by the medical team and the low number of people surveyed (19 participants). A future direction for using EHR in this sporting event would be to collect enough incidents to run analytics. Studying this data will help event organizers to identify factors that may hinder the success of the tournament and take precautions to ensure the safety of athletes, coaches, and the audience.

    Keywords: Electronic health record, sporting venue

  • Brain-Computer Interface for Image Retrieval from EEG

    Abstract: This research is motivated to gift the ability of expression to the speech deprived by building a module that extracts thoughts from the human brain and presents it to the external world as a digital image. This novel approach reconstructs the same images of different colours from the EEG brain waves. Standalone datasets are used to train the model, obtained using the Enobio8 headset. EEG signals are acquired by providing six different shapes of three unique colors as the visual stimuli. After processing these signals, EEG features are extracted using a convolutional Neural Network (CNN). The extracted features alongside noise are passed as an input to conditional Deep Convolutional Generative Adversarial Networks (cDCGANs) and conditional Generative Adversarial Networks (cGANs), which reconstruct the visual stimuli corresponding to the EEG data. The results obtained by the same are recorded and examined. The paper explicitly compares the working and performance of the cGAN and cDCGAN models.

    Keywords: EEG, Image reconstruction, Deep Learning, Brain Media, Generative adversarial networks (GANs).

  • Predicting Oxygen Utilization & Nurse Staffing Needs For SARS-CoV-2

    Abstract: Supplemental oxygen is an essential part of in-hospital care for most patients hospitalized with SARS-CoV-2 pneumonia. This study seeks to identify hospitalized patients who will require advanced oxygen support (high flow oxygen, CPAP, BiPAP, or mechanical ventilation) within 96 hours of admission using a machine learning model. This information can be useful for hospitals to plan for nurse staffing, as patients will consume more resources and will need more staff assistance.

    Data from 302 SARS-CoV-2 patients was used to create a classifier to predict whether or not patients would require advanced oxygen support within 96 hours of admission. Of the 302 cases, 211 were randomly selected to train the model, and 91 were randomly selected for testing. Through a labeled dataset, we performed supervised learning by using a random forest ensemble model which included demographic, clinical comorbidities, vitals, and laboratory values. We used 5-fold cross-validation to evaluate our trained model and employed a majority vote decision across the five trained models in order to produce the final prediction for a given patient. Through the models, we yielded results through sensitivity, specificity, positive predictive value, negative predictive value, and F1 score with the 91 cases of training data. An additional 24 cases were used to test the validity of the ensemble consensus model. Approximately 40% of all patients progressed to require advanced oxygen support 96 hours after the initial presentation. Although the insight gained from the model may not definitively predict the course of an individual patient, this model may help hospital administrators plan for staffing needs with a 48-hour lead time. Patients on high oxygen support require high acuity beds, which have increased nurse-to-patient ratios. Additional samples may increase its statistical significance. Nevertheless, this model demonstrates the potential and viability of using data science to help manage hospital resources.

    Keywords: COVID-19, nurse staffing, resource management, machine learning, oxygen utilization

  • A Research on Force Estimation from EMG with a CNN-based Deep Model

    Abstract:  This paper develops data-driven deep model to estimate muscle contraction force from surface electromyography (EMG) signals. The proposed estimation model is based on deep convolutional neural networks (CNNs). Information of EMG signals from both time and frequency domains has been utilized as input data to two CNN branches. Raw EMG signals are directly input into one branch in order to extract time domain characteristics, while frequency information is fed into the other CNN branch. These two branches are summarized at a concatenation layer, which is followed by three full-connected layers to estimate force levels. End point force at hand, considering the dumbbell curl exercises, has been measured using a testing apparatus, which was designed with a single point load cell. The force data is used as training data of the deep CNN EMG-force model. For validation, different structures of the CNN model are examined with an estimation index of coefficient of determination using EMG and force data of nine subjects

    Keywords: EMG force estimation, CNN-based model, raw EMG signals, frequency information.

  • A Visual Search System Powered by Locality Sensitive Hashing

    Abstract: This paper presents an end-to-end large―scale visual search system. Several challenges were met during its development: dealing with heptamerous image data, indexing large―scale images for massive data updating, training deep learning models for effective feature representation without a lot of manual work, improving the latency and providing accurate results. We assessed our system on a publicly available dataset consisting of flowers commonly available in the UK. This paper describes our implementation in great details as well as our lessons learnt during the building of such a large-scale system. We used locality sensitive hashing to perform the nearest neighbor search and given the data size, we relied on the random projection technique. In order to represent the image in a smaller dimension we used Bit-ResNet as the underlying model given its fine-tuning setup. Extensive experiments show that our system provides satisfactory results.

     Keywords: Additional Key Words and Phrases: datasets, neural networks, visual search, locality sensitive hashing

  • Emergence of Equal Cooperation Induced by Characteristics of Adolescence than Adulthood in an Interrole Conflict Game among Reinforcement Learning Agents

    Abstract: People usually belong to multiple groups, and in such situations “interrole conflicts”occur. Studies on interrole conflicts have mainly targeted subjects after adulthood, although they do not occur only after the developmental stage of adulthood. In addition, limited studies have been conducted on interrole conflicts during adolescence. Furthermore, simulation studies about interrole conflicts have rarely been conducted. The purpose of this study is to clarify the difference between adolescence and adulthood in how to deal with interrole conflict situations. We propose an interrole conflict game (ICG) as a new game-theoretic framework to deal with interrole conflicts and adopt reinforcement learning agents with characteristics of adolescence or adulthood as players. Our multi-agent simulation (MAS) experiment results suggest high learning rate and low discount rate that cause typical adolescent characteristics including risk-taking, impulsivity and novelty seeking can be played important roles to cope an interrole conflict and for emergence of equal cooperation among adolescents in the interrole conflict situation.

    Keywords: interrole conflict game (ICG), reinforcement learning, multi-agent simulation (MAS), characteristics of adolescence or adulthood, emergence of equal cooperation.

  • Drone Guidance System based Simultaneous Localization and Mapping for Free Parking Space Localization

    Abstract:The rapid development of drone has changed the way of our lifestyle by helping us to deal with many issues we were unable to solve before. Such as scanning a large farm area for tracking boars using the drone integrated CMOS camera and so on. In this research, we built a map for drone guidance system in order to detect a fee parking space. The Simultaneous Localization and Mapping (SLAM) is used as a control method. “SLAM”, is the process of mapping an area whilst keeping track of the location of the device within that area. A technology that enables autonomous flight even in an environment where GPS cannot be used . In addition, we used a Versatile and Accurate Monocular SLAM (ORB-SLAM) for real time operation and Large-Scale Direct Monocular Simultaneous Localization and Mapping (LSD-SLAM) which are typical SLAMs based visual SLAM. The experimental result shows that, although the generated map was somehow difficult to visualized but due the camera’s self-position estimation gave a rough route of the path which enable the drone to locate the free parking lots. Our future focus will be to implement the automatic flying system based on the generated map and the improvement of the ORB-SLAM features.

    Keywords: slam; cmos camera ; orb-slam, parking lots; localization, guidance system