Search results for: train accident

Authors: Qianwei Cheng, A. K. M. Mahbubur Rahman, Anis Sarker, Abu Bakar Siddik Nayem, Ovi Paul, Amin Ahsan Ali, M. Ashraful Amin, Ryosuke Shibasaki, Moinul Zaber

Abstract:

Thomas Friedman, in his famous book, argued that the world in this 21st century is flat and will continue to be flatter. This is attributed to rapid globalization and the interdependence of humanity that engendered tremendous in-flow of human migration towards the urban spaces. In order to keep the urban environment sustainable, policy makers need to plan based on extensive analysis of the urban environment. With the advent of high definition satellite images, high resolution data, computational methods such as deep neural network analysis, and hardware capable of high-speed analysis; urban planning is seeing a paradigm shift. Legacy data on urban environments are now being complemented with high-volume, high-frequency data. However, the first step of understanding urban space lies in useful categorization of the space that is usable for data collection, analysis, and visualization. In this paper, we propose a pragmatic categorization method that is readily usable for machine analysis and show applicability of the methodology on a developing world setting. Categorization to plan sustainable urban spaces should encompass the buildings and their surroundings. However, the state-of-the-art is mostly dominated by classification of building structures, building types, etc. and largely represents the developed world. Hence, these methods and models are not sufficient for developing countries such as Bangladesh, where the surrounding environment is crucial for the categorization. Moreover, these categorizations propose small-scale classifications, which give limited information, have poor scalability and are slow to compute in real time. Our proposed method is divided into two steps-categorization and automation. We categorize the urban area in terms of informal and formal spaces and take the surrounding environment into account. 50 km × 50 km Google Earth image of Dhaka, Bangladesh was visually annotated and categorized by an expert and consequently a map was drawn. The categorization is based broadly on two dimensions-the state of urbanization and the architectural form of urban environment. Consequently, the urban space is divided into four categories: 1) highly informal area; 2) moderately informal area; 3) moderately formal area; and 4) highly formal area. In total, sixteen sub-categories were identified. For semantic segmentation and automatic categorization, Google’s DeeplabV3+ model was used. The model uses Atrous convolution operation to analyze different layers of texture and shape. This allows us to enlarge the field of view of the filters to incorporate larger context. Image encompassing 70% of the urban space was used to train the model, and the remaining 30% was used for testing and validation. The model is able to segment with 75% accuracy and 60% Mean Intersection over Union (mIoU). In this paper, we propose a pragmatic categorization method that is readily applicable for automatic use in both developing and developed world context. The method can be augmented for real-time socio-economic comparative analysis among cities. It can be an essential tool for the policy makers to plan future sustainable urban spaces.

Keywords: semantic segmentation, urban environment, deep learning, urban building, classification

Procedia PDF Downloads 191

Authors: Gabrielle Robertson, Matthew Downs, Joseph Dagher

Abstract:

Iron deposition in the brain has been linked with a host of neurological disorders such as Alzheimer’s, Parkinson’s, and Multiple Sclerosis. While some treatment options exist, there are no objective measurement tools that allow for the monitoring of iron levels in the brain in vivo. An emerging Magnetic Resonance Imaging (MRI) method has been recently proposed to deduce iron concentration through quantitative measurement of magnetic susceptibility. This is a multi-step process that involves repeated modeling of physical processes via approximate numerical solutions. For example, the last two steps of this Quantitative Susceptibility Mapping (QSM) method involve I) mapping magnetic field into magnetic susceptibility and II) mapping magnetic susceptibility into iron concentration. Process I involves solving an ill-posed inverse problem by using regularization via injection of prior belief. The end result from Process II highly depends on the model used to describe the molecular content of each voxel (type of iron, water fraction, etc.) Due to these factors, the accuracy and repeatability of QSM have been an active area of research in the MRI and medical imaging community. This work aims to estimate iron concentration in the brain via a single step. A synthetic numerical model of the human head was created by automatically and manually segmenting the human head on a high-resolution grid (640x640x640, 0.4mm³) yielding detailed structures such as microvasculature and subcortical regions as well as bone, soft tissue, Cerebral Spinal Fluid, sinuses, arteries, and eyes. Each segmented region was then assigned tissue properties such as relaxation rates, proton density, electromagnetic tissue properties and iron concentration. These tissue property values were randomly selected from a Probability Distribution Function derived from a thorough literature review. In addition to having unique tissue property values, different synthetic head realizations also possess unique structural geometry created by morphing the boundary regions of different areas within normal physical constraints. This model of the human brain is then used to create synthetic MRI measurements. This is repeated thousands of times, for different head shapes, volume, tissue properties and noise realizations. Collectively, this constitutes a training-set that is similar to in vivo data, but larger than datasets available from clinical measurements. This 3D convolutional U-Net neural network architecture was used to train data-driven Deep Learning models to solve for iron concentrations from raw MRI measurements. The performance was then tested on both synthetic data not used in training as well as real in vivo data. Results showed that the model trained on synthetic MRI measurements is able to directly learn iron concentrations in areas of interest more effectively than other existing QSM reconstruction methods. For comparison, models trained on random geometric shapes (as proposed in the Deep QSM method) are less effective than models trained on realistic synthetic head models. Such an accurate method for the quantitative measurement of iron deposits in the brain would be of important value in clinical studies aiming to understand the role of iron in neurological disease.

Keywords: magnetic resonance imaging, MRI, iron deposition, machine learning, quantitative susceptibility mapping

Procedia PDF Downloads 136

Authors: Ugo Galvanetto, Pirto G. Pavan, Mirco Zaccariotto

Abstract:

The discipline of Computer-integrated surgery (CIS) will provide equipment able to improve the efficiency of healthcare systems and, which is more important, clinical results. Surgeons and machines will cooperate in new ways that will extend surgeons’ ability to train, plan and carry out surgery. Patient specific CIS of the brain requires several steps: 1 - Fast generation of brain models. Based on image recognition of MR images and equipped with artificial intelligence, image recognition techniques should differentiate among all brain tissues and segment them. After that, automatic mesh generation should create the mathematical model of the brain in which the various tissues (white matter, grey matter, cerebrospinal fluid …) are clearly located in the correct positions. 2 – Reliable and fast simulation of the surgical process. Computational mechanics will be the crucial aspect of the entire procedure. New algorithms will be used to simulate the mechanical behaviour of cutting through cerebral tissues. 3 – Real time provision of visual and haptic feedback A sophisticated human-machine interface based on ergonomics and psychology will provide the feedback to the surgeon. The present work will address in particular point 2. Modelling the cutting of soft tissue in a structure as complex as the human brain is an extremely challenging problem in computational mechanics. The finite element method (FEM), that accurately represents complex geometries and accounts for material and geometrical nonlinearities, is the most used computational tool to simulate the mechanical response of soft tissues. However, the main drawback of FEM lies in the mechanics theory on which it is based, classical continuum Mechanics, which assumes matter is a continuum with no discontinuity. FEM must resort to complex tools such as pre-defined cohesive zones, external phase-field variables, and demanding remeshing techniques to include discontinuities. However, all approaches to equip FEM computational methods with the capability to describe material separation, such as interface elements with cohesive zone models, X-FEM, element erosion, phase-field, have some drawbacks that make them unsuitable for surgery simulation. Interface elements require a-priori knowledge of crack paths. The use of XFEM in 3D is cumbersome. Element erosion does not conserve mass. The Phase Field approach adopts a diffusive crack model instead of describing true tissue separation typical of surgical procedures. Modelling discontinuities, so difficult when using computational approaches based on classical continuum Mechanics, is instead easy for novel computational methods based on Peridynamics (PD). PD is a non-local theory of mechanics formulated with no use of spatial derivatives. Its governing equations are valid at points or surfaces of discontinuity, and it is, therefore especially suited to describe crack propagation and fragmentation problems. Moreover, PD does not require any criterium to decide the direction of crack propagation or the conditions for crack branching or coalescence; in the PD-based computational methods, cracks develop spontaneously in the way which is the most convenient from an energy point of view. Therefore, in PD computational methods, crack propagation in 3D is as easy as it is in 2D, with a remarkable advantage with respect to all other computational techniques.

Keywords: computational mechanics, peridynamics, finite element, biomechanics

Procedia PDF Downloads 80

Authors: Utkarsh A. Mishra, Ankit Bansal

Abstract:

At high temperature, radiative heat transfer is the dominant mode of heat transfer. It is governed by various phenomena such as photon emission, absorption, and scattering. The solution of the governing integrodifferential equation of radiative transfer is a complex process, more when the effect of participating medium and wavelength properties are taken into consideration. Although a generic formulation of such radiative transport problem can be modeled for a wide variety of problems with non-gray, non-diffusive surfaces, there is always a trade-off between simplicity and accuracy of the problem. Recently, solutions of complicated mathematical problems with statistical methods based on randomization of naturally occurring phenomena have gained significant importance. Photon bundles with discrete energy can be replicated with random numbers describing the emission, absorption, and scattering processes. Photon Monte Carlo (PMC) is a simple, yet powerful technique, to solve radiative transfer problems in complicated geometries with arbitrary participating medium. The method, on the one hand, increases the accuracy of estimation, and on the other hand, increases the computational cost. The participating media -generally a gas, such as CO₂, CO, and H₂O- present complex emission and absorption spectra. To model the emission/absorption accurately with random numbers requires a weighted sampling as different sections of the spectrum carries different importance. Importance sampling (IS) was implemented to sample random photon of arbitrary wavelength, and the sampled data provided unbiased training of MC estimators for better results. A better replacement to uniform random numbers is using deterministic, quasi-random sequences. Halton, Sobol, and Faure Low-Discrepancy Sequences are used in this study. They possess better space-filling performance than the uniform random number generator and gives rise to a low variance, stable Quasi-Monte Carlo (QMC) estimators with faster convergence. An optimal supervised learning scheme was further considered to reduce the computation costs of the PMC simulation. A one-dimensional plane-parallel slab problem with participating media was formulated. The history of some randomly sampled photon bundles is recorded to train an Artificial Neural Network (ANN), back-propagation model. The flux was calculated using the standard quasi PMC and was considered to be the training target. Results obtained with the proposed model for the one-dimensional problem are compared with the exact analytical and PMC model with the Line by Line (LBL) spectral model. The approximate variance obtained was around 3.14%. Results were analyzed with respect to time and the total flux in both cases. A significant reduction in variance as well a faster rate of convergence was observed in the case of the QMC method over the standard PMC method. However, the results obtained with the ANN method resulted in greater variance (around 25-28%) as compared to the other cases. There is a great scope of machine learning models to help in further reduction of computation cost once trained successfully. Multiple ways of selecting the input data as well as various architectures will be tried such that the concerned environment can be fully addressed to the ANN model. Better results can be achieved in this unexplored domain.

Keywords: radiative heat transfer, Monte Carlo Method, pseudo-random numbers, low discrepancy sequences, artificial neural networks

Procedia PDF Downloads 223

Authors: Mian Huang, Chi Ma, Junyu Lin, William Lu

Abstract:

Introduction: Bone health attracts more attention recently and an intelligent question and answer (QA) chatbot for osteoporosis is helpful for science popularization. With Generative Pretrained Transformer (GPT) technology developing, we build an osteoporosis corpus dataset and then fine-tune LLaMA, a famous open-source GPT foundation large language model(LLM), on our self-constructed osteoporosis corpus. Evaluated by clinical orthopedic experts, our fine-tuned model outperforms vanilla LLaMA on osteoporosis QA task in Chinese. Three-dimensional quantitative computed tomography (QCT) measured bone mineral density (BMD) is considered as more accurate than DXA for BMD measurement in recent years. We develop an automatic Phantom-less QCT(PL-QCT) that is more efficient for BMD measurement since no need of an external phantom for calibration. Combined with LLM on osteoporosis, our PL-QCT provides efficient and accurate BMD measurement for our chatbot users. Material and Methods: We build an osteoporosis corpus containing about 30,000 Chinese literatures whose titles are related to osteoporosis. The whole process is done automatically, including crawling literatures in .pdf format, localizing text/figure/table region by layout segmentation algorithm and recognizing text by OCR algorithm. We train our model by continuous pre-training with Low-rank Adaptation (LoRA, rank=10) technology to adapt LLaMA-7B model to osteoporosis domain, whose basic principle is to mask the next word in the text and make the model predict that word. The loss function is defined as cross-entropy between the predicted and ground-truth word. Experiment is implemented on single NVIDIA A800 GPU for 15 days. Our automatic PL-QCT BMD measurement adopt AI-associated region-of-interest (ROI) generation algorithm for localizing vertebrae-parallel cylinder in cancellous bone. Due to no phantom for BMD calibration, we calculate ROI BMD by CT-BMD of personal muscle and fat. Results & Discussion: Clinical orthopaedic experts are invited to design 5 osteoporosis questions in Chinese, evaluating performance of vanilla LLaMA and our fine-tuned model. Our model outperforms LLaMA on over 80% of these questions, understanding ‘Expert Consensus on Osteoporosis’, ‘QCT for osteoporosis diagnosis’ and ‘Effect of age on osteoporosis’. Detailed results are shown in appendix. Future work may be done by training a larger LLM on the whole orthopaedics with more high-quality domain data, or a multi-modal GPT combining and understanding X-ray and medical text for orthopaedic computer-aided-diagnosis. However, GPT model gives unexpected outputs sometimes, such as repetitive text or seemingly normal but wrong answer (called ‘hallucination’). Even though GPT give correct answers, it cannot be considered as valid clinical diagnoses instead of clinical doctors. The PL-QCT BMD system provided by Bone’s QCT(Bone’s Technology(Shenzhen) Limited) achieves 0.1448mg/cm2(spine) and 0.0002 mg/cm2(hip) mean absolute error(MAE) and linear correlation coefficient R2=0.9970(spine) and R2=0.9991(hip)(compared to QCT-Pro(Mindways)) on 155 patients in three-center clinical trial in Guangzhou, China. Conclusion: This study builds a Chinese osteoporosis corpus and develops a fine-tuned and domain-adapted LLM as well as a PL-QCT BMD measurement system. Our fine-tuned GPT model shows better capability than LLaMA model on most testing questions on osteoporosis. Combined with our PL-QCT BMD system, we are looking forward to providing science popularization and early morning screening for potential osteoporotic patients.

Keywords: GPT, phantom-less QCT, large language model, osteoporosis

Procedia PDF Downloads 71

Authors: Subhash Nerella, Ziyuan Guan, Azra Bihorac, Parisa Rashidi

Abstract:

Intensive care units (ICUs) provide care to critically ill patients in severe and life-threatening conditions. However, patient monitoring in the ICU is limited by the time and resource constraints imposed on healthcare providers. Many critical care indices such as mobility are still manually assessed, which can be subjective, prone to human errors, and lack granularity. Other important aspects, such as environmental factors, are not monitored at all. For example, critically ill patients often experience circadian disruptions due to the absence of effective environmental “timekeepers” such as the light/dark cycle and the systemic effect of acute illness on chronobiologic markers. Although the occurrence of delirium is associated with circadian disruption risk factors, these factors are not routinely monitored in the ICU. Hence, there is a critical unmet need to develop systems for precise and real-time assessment through novel enabling technologies. We have developed the mobility and circadian disruption quantification system (Mobi-DiQ) by augmenting biomarker and clinical data with pervasive sensing data to generate mobility and circadian cues related to mobility, nightly disruptions, and light and noise exposure. We hypothesize that Mobi-DiQ can provide accurate mobility and circadian cues that correlate with bedside clinical mobility assessments and circadian biomarkers, ultimately important for delirium risk assessment and prevention. The collected multimodal dataset consists of depth images, Electromyography (EMG) data, patient extremity movement captured by accelerometers, ambient light levels, Sound Pressure Level (SPL), and indoor air quality measured by volatile organic compounds, and the equivalent CO₂ concentration. For delirium risk assessment, the system recognizes mobility cues (axial body movement features and body key points) and circadian cues, including nightly disruptions, ambient SPL, and light intensity, as well as other environmental factors such as indoor air quality. The Mobi-DiQ system consists of three major components: the pervasive sensing system, a data storage and analysis server, and a data annotation system. For data collection, six local pervasive sensing systems were deployed, including a local computer and sensors. A video recording tool with graphical user interface (GUI) developed in python was used to capture depth image frames for analyzing patient mobility. All sensor data is encrypted, then automatically uploaded to the Mobi-DiQ server through a secured VPN connection. Several data pipelines are developed to automate the data transfer, curation, and data preparation for annotation and model training. The data curation and post-processing are performed on the server. A custom secure annotation tool with GUI was developed to annotate depth activity data. The annotation tool is linked to the MongoDB database to record the data annotation and to provide summarization. Docker containers are also utilized to manage services and pipelines running on the server in an isolated manner. The processed clinical data and annotations are used to train and develop real-time pervasive sensing systems to augment clinical decision-making and promote targeted interventions. In the future, we intend to evaluate our system as a clinical implementation trial, as well as to refine and validate it by using other data sources, including neurological data obtained through continuous electroencephalography (EEG).

Keywords: deep learning, delirium, healthcare, pervasive sensing

Procedia PDF Downloads 93

Authors: Pei Yi Lin

Abstract:

Objective: The objective of this study is to use machine learning methods to build an early prediction classifier model for acute delirium to improve the quality of medical care for intensive care patients. Background: Delirium is a common acute and sudden disturbance of consciousness in critically ill patients. After the occurrence, it is easy to prolong the length of hospital stay and increase medical costs and mortality. In 2021, the incidence of delirium in the intensive care unit of internal medicine was as high as 59.78%, which indirectly prolonged the average length of hospital stay by 8.28 days, and the mortality rate is about 2.22% in the past three years. Therefore, it is expected to build a delirium prediction classifier through big data analysis and machine learning methods to detect delirium early. Method: This study is a retrospective study, using the artificial intelligence big data database to extract the characteristic factors related to delirium in intensive care unit patients and let the machine learn. The study included patients aged over 20 years old who were admitted to the intensive care unit between May 1, 2022, and December 31, 2022, excluding GCS assessment <4 points, admission to ICU for less than 24 hours, and CAM-ICU evaluation. The CAMICU delirium assessment results every 8 hours within 30 days of hospitalization are regarded as an event, and the cumulative data from ICU admission to the prediction time point are extracted to predict the possibility of delirium occurring in the next 8 hours, and collect a total of 63,754 research case data, extract 12 feature selections to train the model, including age, sex, average ICU stay hours, visual and auditory abnormalities, RASS assessment score, APACHE-II Score score, number of invasive catheters indwelling, restraint and sedative and hypnotic drugs. Through feature data cleaning, processing and KNN interpolation method supplementation, a total of 54595 research case events were extracted to provide machine learning model analysis, using the research events from May 01 to November 30, 2022, as the model training data, 80% of which is the training set for model training, and 20% for the internal verification of the verification set, and then from December 01 to December 2022 The CU research event on the 31st is an external verification set data, and finally the model inference and performance evaluation are performed, and then the model has trained again by adjusting the model parameters. Results: In this study, XG Boost, Random Forest, Logistic Regression, and Decision Tree were used to analyze and compare four machine learning models. The average accuracy rate of internal verification was highest in Random Forest (AUC=0.86), and the average accuracy rate of external verification was in Random Forest and XG Boost was the highest, AUC was 0.86, and the average accuracy of cross-validation was the highest in Random Forest (ACC=0.77). Conclusion: Clinically, medical staff usually conduct CAM-ICU assessments at the bedside of critically ill patients in clinical practice, but there is a lack of machine learning classification methods to assist ICU patients in real-time assessment, resulting in the inability to provide more objective and continuous monitoring data to assist Clinical staff can more accurately identify and predict the occurrence of delirium in patients. It is hoped that the development and construction of predictive models through machine learning can predict delirium early and immediately, make clinical decisions at the best time, and cooperate with PADIS delirium care measures to provide individualized non-drug interventional care measures to maintain patient safety, and then Improve the quality of care.

Keywords: critically ill patients, machine learning methods, delirium prediction, classifier model

Procedia PDF Downloads 75