Proposed Machine Learning Framework Research Articles

Acoustic emission signatures for quantifying damage patterns in half grouted sleeve connection under tensile load

The half-grouted sleeve connection (HGSC) have been widely used as connection tools for prefabricated concrete (PC) buildings. It is crucial to maintain the safety and stability of the nodes and detecting internal defects in the grouting sleeve for the safety of the entire structure. Therefore, understanding the mechanical behavior and damage mechanism of HGSC is essential for the design and construction of PC. Currently, there are few on-site techniques applicable for monitoring the damage progress. The acoustic emission (AE) as a passive method shows great potential to characterize the damage in the HGSC. Therefore, in this paper, we propose to use AE to quantify the damage of the HGSC with defects under the tensile load. The artificial defects were introduced into HGSC with different defective rates. The monotonic tensile load was applied to each HGSC sample with a loading rate of 0.33 mm/s and tested to the failure. In this process, the AE signals were collected by AE transducers. For the classification and prediction of damage signals, we utilized a machine learning framework tailored for HGSC. This framework integrates unsupervised k-means++ cluster analysis and supervised K-Nearest Neighbors (K-NN) for damage classification and prediction. The experimental results showed that four AE signal clusters can be defined during the progress of tensile failure:(i) microcrack expansion;(ii) the friction signals of rebar and high-strength grouts; (iii)the generation of microcracks; (iv) plastic deformation of rebar. The corresponding AE signatures were identified. The K-NN model trained on the training data provides a high accuracy in predicting different class of defects. It was concluded that the proposed machine learning framework is a promising and powerful tool for realizing the on-site damage detection and prediction for HGSC.

Examining the dynamic engagement process of passengers in online car-hailing system: a view of user value

ABSTRACT The online car-hailing system developed rapidly in recent years. This paper introduces the concept of user value and develops a machine learning framework for modeling the dynamic transformation of passenger value in online car-hailing system. Specifically, during the engagement process, four user value types representing the engagement status were captured based on both consumption and travel attributes. Additionally, we modeled user value conversion prediction throughout the engagement process, seamlessly integrating the benefits of features enhancement and time series prediction. Experiments using real order data demonstrated that the proposed machine learning framework has higher predictive accuracy than other methods. The value categories of online car-hailing users change over time. High-value users have high retention rates, while low-value users are the opposite, and medium-value users have the greatest potential. The results provide an important basis for modeling the engagement process of online car-hailing passengers, and for promoting the customer maintenance and management in online car-hailing systems.

A tensile properties-related fatigue strength predicted machine learning framework for alloys used in aerospace

A tensile properties-related fatigue strength prediction framework based on machine learning (ML) methods was proposed. Firstly, 200 data containing six materials used in aerospace were collected. Secondly, tensile properties-related ML framework, including residual neural network (ResNet) model, gradient boosting decision tree (GBDT) model and LightGBM (LGB) model, was established to predict fatigue strength under symmetric cycles (R = -1) after features correlation analysis. Then, Gaussian noise was introduced into the features to prove the robustness and feature sensitivity of proposed ML framework. Finally, material characteristics of fatigue resistance in the next generation were discussed after the analysis of the correlation between the tensile properties and fatigue strength. Compared with ± 50 % error band of classical prediction models, the prediction results of proposed ML framework located in the ± 10 % error band, especially for the special crystal orientations and stress ratios. In practical applications, the tensile properties-related ML framework could be applied efficiently to materials fatigue strength prediction used in aerospace due to its robust prediction accuracy and generalization ability without additional parameters fitting.

Abstract 878: Enhancing single-cell RNA sequencing analysis in cancer research: A machine learning framework based on LightGBM for automated cell type annotation

Abstract Single-cell RNA sequencing (scRNA-seq) has been widely used in cancer research to understand the complex gene expression diversity and cancer heterogeneity. However, manual annotation of cell types in the scRNA-seq pipeline is time-consuming and depends on the expertise of analyzers, which can significantly influence the results of downstream analyses. To address this problem, we proposed a novel machine learning framework utilizing the LightGBM model for automated and efficient cell-type annotation of scRNA-seq. Two independent scRNA-seq datasets of non-small cell lung cancer (NSCLC) downloaded from the Gene Expression Omnibus (GEO) were used to train and test our model. A standard procedure is applied to both scRNA-seq datasets for quality control and preprocessing, in which poor-quality cells with low gene expressions or high scores for cellular stress/death were excluded. In addition, Harmony is applied to mitigate batch effects in scRNA-seq that could cause variability due to non-biological factors in experiments. Nine different cell types, including endothelial, epithelial, fibroblast, macrophages, mast, plasma, pulmonary alveolar, B, and T cells, were manually labeled in the two datasets by the providers, which were also examined using gene markers corresponding to different cell types from PanglaoDB and DAVID. These manually labeled cell types were used as the ground truth for training and testing our model. In the training stage, the training dataset (containing 85,000 cells from 44 NSCLC samples) of scRNA-seq was used to train the LightGBM model with its high-variable genes. Then, the model would be evaluated using an independent test dataset (containing 8,000 cells from 18 NSCLC samples) by comparing the automatically predicted and manually labeled cell types. The training result showed that our model could successfully specify the nine different cell types, achieving an overall average accuracy, F1 score, and precision of 0.86 each respectively. In the independent dataset test, the model demonstrated good generalization, showing high predictive performance across all cell types, with an average accuracy, F1 score, and precision of 0.8, 0.78, and 0.8, respectively. Specific to the predictions in the test dataset, we found that some epithelial cells were mistakenly identified as other cell types. This might be because of the complex gene expression patterns exhibited by tumor epithelial cells, making accurate predictions challenging. The proposed machine learning framework facilitates cell labeling and unravels the intricate heterogeneity within lung cancer datasets. The combination of LightGBM and standardized preprocessing establishes a benchmark for high-throughput, accurate single-cell analysis, paving the way for discoveries that are more targeted and have significant clinical impact. Citation Format: Tsung Hsien Chuang, Liang-Chuan Lai, Tzu-Pin Lu, Mong-Hsun Tsai, Hsiang-Han Chen, Eric Y. Chuang. Enhancing single-cell RNA sequencing analysis in cancer research: A machine learning framework based on LightGBM for automated cell type annotation [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 878.

Enhancement of multirotor UAV conceptual design through Machine Learning algorithms

Designing an efficient and optimized multirotor UAV requires laborious trade-off analyses, involving numerous design variables and mission requirement parameters, especially during the early conceptual design phase. The large number of unknown parameters, as well as the associated design effort often leads to non-optimal designs, for the sake of time efficiency. This work presents the implementation of a machine learning (ML) framework to assist and expedite the conceptual design phase of multirotor UAVs. The framework utilizes information from a comprehensive database of commercial lightweight multirotor UAVs. The database contains an extensive collection of crucial sizing parameters, performance metrics, and features associated with foldability and indoor guidance (e.g., obstacle avoidance sensors). These attributes specifically pertain to multirotor UAVs weighing less than 2kg, which exhibit diverse design and performance characteristics. The proposed ML framework employs multiple regression models (e.g. k-nearest neighbors regression, multi-layer perceptron regression) to predict the sizing parameters during a multirotor UAV’s conceptual design phase. This enables designers to make quick informed decisions, while also significantly reducing computational time and effort. Finally, the ML framework’s predictive capability is validated by comparing the predicted values with real-world data from an “unseen” test dataset.

TO‐NODE: Topology optimization with neural ordinary differential equation

AbstractDeep learning methods have become attractive recently to accelerate topology optimization (TO) because of their capability to save huge computational costs with negligible sacrifice in the quality of final topologies. However, most current approaches rely on numerical mechanics platforms for training which incur significant computation costs. Further, current approaches are not suited for dynamically changing boundary conditions. This is because these deep learning frameworks once trained using results generated with a specific set of boundary conditions do not readily adapt to others. In this article, we present a novel approach to leverage the abilities of deep learning models for TO. We demonstrate that optimization can be achieved using a neural ordinary differential equation. The evolution of the design variable in each iteration of TO is then achieved by numerical integration of this neural differential equation with respect to the starting design. To improve the quality of the results, two levels of generative adversarial networks are also introduced, at the sequence, and image levels, respectively. The proposed machine learning framework is capable of generating full optimization paths relevant to TO in high resolution within seconds and can address novel unseen boundary conditions.

A general framework of high-performance machine learning algorithms: application in structural mechanics

Data-driven models utilizing powerful artificial intelligence (AI) algorithms have been implemented over the past two decades in different fields of simulation-based engineering science. Most numerical procedures involve processing data sets developed from physical or numerical experiments to create closed-form formulae to predict the corresponding systems’ mechanical response. Efficient AI methodologies that will allow the development and use of accurate predictive models for solving computational intensive engineering problems remain an open issue. In this research work, high-performance machine learning (ML) algorithms are proposed for modeling structural mechanics-related problems, which are implemented in parallel and distributed computing environments to address extremely computationally demanding problems. Four machine learning algorithms are proposed in this work and their performance is investigated in three different structural engineering problems. According to the parametric investigation of the prediction accuracy, the extreme gradient boosting with extended hyper-parameter optimization (XGBoost-HYT-CV) was found to be more efficient regarding the generalization errors deriving a 4.54% residual error for all test cases considered. Furthermore, a comprehensive statistical analysis of the residual errors and a sensitivity analysis of the predictors concerning the target variable are reported. Overall, the proposed models were found to outperform the existing ML methods, where in one case the residual error was decreased by 3-fold. Furthermore, the proposed algorithms demonstrated the generic characteristic of the proposed ML framework for structural mechanics problems.

Application of artificial neural networks to predict the particle-scale contact force of photoelastic disks

Measuring the contact force of photoelastic granular materials in a two-dimensional system either has a high dependence on the experimental parameters or requires expensive computations. This study proposes a supervised machine learning strategy that constructs a universal relation between the features of a stressed photoelastic disk and the corresponding total contact force acting on it via an artificial neural network (ANN). Most of the common circumstances of particle–particle contact that occurs in real photoelastic granular flow experiments are considered in the data. The ANN reached a test R-squared value of about 0.96. Therefore, the proposed machine learning framework is an efficient and accurate force measurement technique for two-dimensional photoelastic granular flows.

Machine Learning for Film Thickness Prediction in Elastohydrodynamic Lubricated Elliptical Contacts

This study extends the use of Machine Learning (ML) approaches for lubricant film thickness predictions to the general case of elliptical elastohydrodynamic (EHD) contacts, by considering wide and narrow contacts over a wide range of ellipticity and operating conditions. Finite element (FEM) simulations are used to generate substantial training and testing datasets that are used within the proposed ML framework. The complete dataset entails 915 samples; split into an 823-sample training dataset and a 92-sample testing dataset, corresponding to 90% and 10% of the combined dataset samples, respectively. The proposed ML model consists of a pre-processing stage in which conventional EHD dimensionless groups are used to minimize the number of inputs into the model, reducing them to only three. The core of the model is based on Gaussian Process Regression (GPR), a powerful ML regression tool, well-suited for small-sized datasets, producing output central and minimum film thicknesses, also in dimensionless form. The last stage is a post-processing one, in which the output film thicknesses are retrieved in dimensional from. The results reveal the capabilities and potential of the proposed ML framework, producing quasi-instantaneous predictions that are far more accurate than conventional film thickness analytical formulae. In fact, the produced central and minimum film thickness predictions are on average within 0.3% and 1.0% of the FEM results, respectively.

CAD system for epileptic seizure detection from EEG through image processing and SURF-BOF technique

Abstract Epilepsy is one of the most debilitating neurological diseases that abruptly alters a person’s way of life. Manual diagnosis is a laborious and time-consuming task prone to human error. Therefore, automating this task by developing an intelligent system is necessary. Existing deep learning (DL) models require high training time, large datasets, and machines with more memory and processing power. In addition, owing to the black-box nature of DL models, no one can determine the features that the network prefers for classification decisions. To overcome these challenges, this study proposes an accurate, automatic, and fast-intelligent system for epilepsy detection using a computer-aided diagnosis (CAD) -two-dimensional machine learning (ML) framework. Existing ML models struggle to produce reliable and acceptable diagnostic results owing to the low amplitude and nonstationary nature of electroencephalograms (EEGs), particularly in clinical situations where environmental influences are almost impossible to eliminate. The proposed model was built using the Children’s Hospital Boston and the Massachusetts Institute of Technology dataset, and represents the first study that employs the speeded-up robust feature (SURF) bag of features technique for this application, which generates local features from spectrogram images of the respective one-dimensional EEG signal inputs. In addition, DL features were extracted from the spectrogram images for model performance comparison. Both features were used separately to train the ML classifiers. Implementing SURF offers fast computation and makes the model invariant to distortions, noise, scaling, and so on. Therefore, the proposed model is more suitable for real-time applications, and this ML framework provides an enhanced accuracy of 99.78% compared to the support vector machine-RBF classifier, along with 99.56% sensitivity, 100% specificity, and an error rate of 0.22%. The higher detection accuracy demonstrates the effectiveness of the proposed framework for medical disease diagnosis applications.

Multiscale monitoring of industrial chemical process using wavelet-entropy aided machine learning approach

In recent decades, machine learning (ML) techniques have been effectively applied for industrial process monitoring to assure safety and high-quality yield. Traditional process fault detection, identification, and diagnosis (FDI&D) approaches are insufficiently smart to address the modern complex challenges of real-time industrial chemical processes. The detection and diagnostic resolution of the traditional monitoring approach are less robust, inefficient, and produce the wrong interpretation of actual fault information. These approaches are based on a single-scale fault illustration and cannot effectively address multiple fault depiction roots. This study introduces a novel ML-aided methodological framework for industrial and manufacturing monitoring systems. The proposed ML framework is developed using Principal component analysis (PCA), Shannon information entropy (IE), wavelet transformation (WT), and signed directed graph (SDG). It includes fault detection, identification, and diagnostic propagation root-path interpretation to address the safety challenges of modern, real-time industrial chemical processes. The proposed methodological framework is validated using the Tennessee Eastman process (TEP) benchmark to highlight their performance and efficiency. The results of this study determined that the new proposed approach is more efficient in terms of accuracy, robustness, and actual propagation root cause than traditional techniques. It has a high fault detection rate (FDR), low fault alarm rate (FAR) that identifies and recognizes the actual faulty-correlated variables to establish the SDG model framework for determining the actual diagnostic propagation root path. It initially enables operators to react to unusual incidents, ensuring industrial safety, minimizing economic loss, and avoiding disasters.

A novel optimal wavelet filter banks for automated diagnosis of Alzheimer’s disease and mild cognitive impairment using Electroencephalogram signals

Electroencephalogram (EEG) of Alzheimer’s disease (AD) patients show a slowing effect and less synchronization. EEG signal’s transient and abrupt nature is captured from various mother wavelets. However, better performance can be obtained by balancing time–frequency localization in wavelet filters. We propose a new approach for designing filter banks based on optimal time–frequency localization of a four-step lifting structure (FSLS). First, we design a FSLS using Euler’s Frobenius half-band polynomial (EFHBP). The perfect reconstruction condition and vanishing moments are imposed on EFHBP to achieve maximum flat half-band filters (HBFs). HBFs are optimized using a balanced uncertainty metric to obtain a time–frequency spread balance. Afterward, these HBFs are used in the FSLS to obtain the synthesis and analysis filter banks. The proposed time–frequency optimized biorthogonal wavelet filter banks (TFOBWFBs) achieved a balance between time–frequency localization. Further, these TFOBWFBs are applied to decompose the EEG signals of AD patients. Twenty different features were extracted from decomposed EEG subbands, of which twelve significant features were selected using Kruskal Walli’s test. The machine learning models were trained and tested using obtained features with 10-fold cross-validation and leave-one-subject-out cross-validation. To validate this study, the proposed TFOBWFBs have been applied to two publicly available EEG datasets of mild cognitive impairment (MCI), AD, and healthy control (HC) subjects. The proposed TFOBWFBs achieved 98.90% accuracy for 2-way (AD vs. HC) and 96.50% for 3-way (AD vs. MCI vs. HC) classification using a support vector machine model with a 10-fold cross-validation. The proposed method outperforms the existing AD and MCI detection techniques. The performance of the proposed machine learning framework with optimization of wavelet filter banks is more significant and fast compared to previous studies. Also, the proposed framework can be applied to detect other neurodegenerative disorders.

Adaptive physics-informed neural operator for coarse-grained non-equilibrium flows

This work proposes a new machine learning (ML)-based paradigm aiming to enhance the computational efficiency of non-equilibrium reacting flow simulations while ensuring compliance with the underlying physics. The framework combines dimensionality reduction and neural operators through a hierarchical and adaptive deep learning strategy to learn the solution of multi-scale coarse-grained governing equations for chemical kinetics. The proposed surrogate’s architecture is structured as a tree, with leaf nodes representing separate neural operator blocks where physics is embedded in the form of multiple soft and hard constraints. The hierarchical attribute has two advantages: (i) It allows the simplification of the training phase via transfer learning, starting from the slowest temporal scales; (ii) It accelerates the prediction step by enabling adaptivity as the surrogate’s evaluation is limited to the necessary leaf nodes based on the local degree of non-equilibrium of the gas. The model is applied to the study of chemical kinetics relevant for application to hypersonic flight, and it is tested here on pure oxygen gas mixtures. In 0-textrm{D} scenarios, the proposed ML framework can adaptively predict the dynamics of almost thirty species with a maximum relative error of 4.5% for a wide range of initial conditions. Furthermore, when employed in 1-textrm{D} shock simulations, the approach shows accuracy ranging from 1% to 4.5% and a speedup of one order of magnitude compared to conventional implicit schemes employed in an operator-splitting integration framework. Given the results presented in the paper, this work lays the foundation for constructing an efficient ML-based surrogate coupled with reactive Navier-Stokes solvers for accurately characterizing non-equilibrium phenomena in multi-dimensional computational fluid dynamics simulations.

Paper-based multiplexed colorimetric biosensing of cardiac and lipid biomarkers integrated with machine learning for accurate acute myocardial infarction early diagnosis and prognosis

This study demonstrates how a colorimetric biosensor based on microfluidic paper can swiftly diagnose a disease and predict its prognosis to triage patients effectively. This was the first biosensor to quantify the gold standard cardiac troponin (cTnI) and lipid biomarkers, including high-density lipoprotein (HDL) and low-density lipoprotein (LDL) simultaneously. Prior research encountered obstacles or limitations, such as measuring a single biomarker or total cholesterol, which cannot distinguish between LDL and HDL. CatBoost, an advanced machine learning (ML) technique used for diagnosis that combines the predictive power of ML algorithms obtained an impressive area under the receiver operating curves (AUROC) of 0.97 ± 0.018 for all possible classification thresholds. CatBoost is a brand-new ensemble framework based on the interaction of multiple health parameters that can generate AUROC values of 0.897 ± 0.047 for the accurate prognosis of recurrent acute myocardial infarction (AMI), demonstrating remarkably accurate AMI diagnosis and prognosis. In addition, this paper-based analytical device (µPAD) biosensor employs an electrophoretic method to overcome the challenges posed by non-specific adsorption. This is accomplished by isolating non-specific biomolecules based on differences in their isoelectric points and removing non-specifically adsorbing colorimetric markers. The limits of detection (LoD) for cTnI in AMI were lower than their respective clinical cutoff values. This study also demonstrated that the proposed ML framework produced significantly better results than conventional statistical analysis. High correlation filtering and (t-SNE) dimensionality reduction were utilized for a limited number of data points. The respectable accuracy and AUROC of this method were also validated using cross-validation.

A machine learning framework for performing binary classification on tabular biomedical data

AbstractBackground and aimOver the past decades, we have witnessed an immense expansion in the arsenal and performance of machine learning (ML) algorithms. One of the most important fields that could benefit from these advancements is biomedical science. To streamline the training and evaluation of binary classifiers, we constructed a universal and flexible ML framework that uses tabular biomedical data as input.Methods and resultsOur framework requires the input data to be provided as a comma-separated values file, in which rows correspond to subjects and columns represent different features. After reading the content of this file, the framework enables the users to perform outlier detection, handle missing values, rescale features, and tackle class imbalance. Then, hyperparameter tuning, feature selection, and internal validation are performed using nested cross-validation. If an additional dataset is available, the framework also provides the option for external validation. Users may also compute SHapley Additive exPlanations values to interpret the individual predictions of the model and identify the most important features. Our ML framework was implemented in Python (version 3.9), and its source code is freely available via GitHub. In the second part of this paper, we also demonstrate the usage of the framework through a case study from the field of cardiovascular imaging.ConclusionsThe proposed ML framework enables the efficient training and evaluation of binary classifiers on tabular biomedical data. We hope our framework will serve as a useful resource for both learning and research purposes and will promote further innovation.

An enhanced Runge Kutta boosted machine learning framework for medical diagnosis

With the development and maturity of machine learning methods, medical diagnosis aided with machine learning methods has become a popular method to assist doctors in diagnosing and treating patients. However, machine learning methods are greatly affected by their hyperparameters, for instance, the kernel parameter in kernel extreme learning machine (KELM) and the learning rate in residual neural networks (ResNet). If the hyperparameters are appropriately set, the performance of the classifier can be significantly improved. To boost the performance of the machine learning methods, this paper proposes to improve the Runge Kutta optimizer (RUN) to adaptively adjust the hyperparameters of the machine learning methods for medical diagnosis purposes. Although RUN has a solid mathematical theoretical foundation, there are still some performance defects when dealing with complex optimization problems. To remedy these defects, this paper proposes a new enhanced RUN method with a grey wolf mechanism and an orthogonal learning mechanism called GORUN. The superior performance of the GORUN was validated against other well-established optimizers on IEEE CEC 2017 benchmark functions. Then, the proposed GORUN is employed to optimize the machine learning models, including the KELM and ResNet, to construct robust models for medical diagnosis. The performance of the proposed machine learning framework was validated on several medical data sets, and the experimental results have demonstrated its superiority.

A voting-based machine learning approach for classifying biological and clinical datasets

BackgroundDifferent machine learning techniques have been proposed to classify a wide range of biological/clinical data. Given the practicability of these approaches accordingly, various software packages have been also designed and developed. However, the existing methods suffer from several limitations such as overfitting on a specific dataset, ignoring the feature selection concept in the preprocessing step, and losing their performance on large-size datasets. To tackle the mentioned restrictions, in this study, we introduced a machine learning framework consisting of two main steps. First, our previously suggested optimization algorithm (Trader) was extended to select a near-optimal subset of features/genes. Second, a voting-based framework was proposed to classify the biological/clinical data with high accuracy. To evaluate the efficiency of the proposed method, it was applied to 13 biological/clinical datasets, and the outcomes were comprehensively compared with the prior methods.ResultsThe results demonstrated that the Trader algorithm could select a near-optimal subset of features with a significant level of p-value < 0.01 relative to the compared algorithms. Additionally, on the large-sie datasets, the proposed machine learning framework improved prior studies by ~ 10% in terms of the mean values associated with fivefold cross-validation of accuracy, precision, recall, specificity, and F-measure.ConclusionBased on the obtained results, it can be concluded that a proper configuration of efficient algorithms and methods can increase the prediction power of machine learning approaches and help researchers in designing practical diagnosis health care systems and offering effective treatment plans.

Machine learning and polymer self-consistent field theory in two spatial dimensions.

A computational framework that leverages data from self-consistent field theory simulations with deep learning to accelerate the exploration of parameter space for block copolymers is presented. This is a substantial two-dimensional extension of the framework introduced in the work of Xuan et al. [J. Comput. Phys. 443, 110519 (2021)]. Several innovations and improvements are proposed. (1) A Sobolev space-trained, convolutional neural network is employed to handle the exponential dimension increase of the discretized, local average monomer density fields and to strongly enforce both spatial translation and rotation invariance of the predicted, field-theoretic intensive Hamiltonian. (2) A generative adversarial network (GAN) is introduced to efficiently and accurately predict saddle point, local average monomer density fields without resorting to gradient descent methods that employ the training set. This GAN approach yields important savings of both memory and computational cost. (3) The proposed machine learning framework is successfully applied to 2D cell size optimization as a clear illustration of its broad potential to accelerate the exploration of parameter space for discovering polymer nanostructures. Extensions to three-dimensional phase discovery appear to be feasible.

MP09-06 ASSESSMENT OF A NOVEL BPMRI-BASED MACHINE LEARNING FRAMEWORK TO AUTOMATE THE DETECTION OF CLINICALLY SIGNIFICANT PROSTATE CANCER USING THE PI-CAI (PROSTATE IMAGING: CANCER AI) CHALLENGE DATASET

You have accessJournal of UrologyCME1 Apr 2023MP09-06 ASSESSMENT OF A NOVEL BPMRI-BASED MACHINE LEARNING FRAMEWORK TO AUTOMATE THE DETECTION OF CLINICALLY SIGNIFICANT PROSTATE CANCER USING THE PI-CAI (PROSTATE IMAGING: CANCER AI) CHALLENGE DATASET Andre Luis Abreu, Giovanni Cacciamani, Masatomo Kaneko, Vasileios Magouliantis, Yijing Yang, Vinay Duddalwar, C-C Jay Kuo, Inderbir Gill, and Chrysostomos L. Nikias Andre Luis AbreuAndre Luis Abreu More articles by this author , Giovanni CacciamaniGiovanni Cacciamani More articles by this author , Masatomo KanekoMasatomo Kaneko More articles by this author , Vasileios MagouliantisVasileios Magouliantis More articles by this author , Yijing YangYijing Yang More articles by this author , Vinay DuddalwarVinay Duddalwar More articles by this author , C-C Jay KuoC-C Jay Kuo More articles by this author , Inderbir GillInderbir Gill More articles by this author , and Chrysostomos L. NikiasChrysostomos L. Nikias More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000003224.06AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: To develop a novel biparametric magnetic resonance imaging (bpMRI)-based machine learning (ML) framework to automatically detect prostate lesions using the PI-CAI (Prostate Imaging: Cancer AI) Challenge Dataset. PI-CAI is an annotated multi-center dataset of 1300 bpMRIs publicly available to validate ML algorithms performance at csPCa (Grade Group ≥2 ) detection and diagnosis. METHODS: We used the Green Learning paradigm for feature extraction, which offers a lightweight model size and explainable feature extraction process. We use the IPHOP-II method that decomposes the input into a spatial-spectral representation, where the discriminant dimensions are kept through feature selection and eventually fed to the classifier. IPHOP-II takes as input T2W, ADC and DWI high b-value to generate per voxel features for classification. The classifier is trained to discern between csPCa and non-csPCa. After training, the ML detector was tested to predict voxel-wise probabilities for csPCa on 100 bpMRIs (unseen). Per-lesion Average Precision metric (AP) and average per-Patient ROC (Area Under the Curve - Receiver Operating Characteristic - AUROC) for the presence of csPCa are used to evaluate our detector. The PI-CAI score is the average of AP and AUROC. Results were also stratified accordingly to PSA density. RESULTS: A total of 1200 bpMRI were used for training the ML model and tested on 100 bpMRIs (Figure 1). In the testing set, the average AP for csPCa detection per lesion was 0.47, with a sensitivity of 0.61 precision 0.42 rate at Youden’s index cutoff. On the patient-level analysis, the AUROC was 0.81 and a sensitivity of 0.92 at a 0.43 false positive rate. The PI-CAI score was 0.64. In the subset analysis considering only patients with PSA density ≥0.2ng/ml2 we found an improvement in the classification task with the AUROC was 0.83 and a sensitivity of 0.94 at a lower 0.41 false positive rate. CONCLUSIONS: These results show the feasibility of a novel bpMRI-based ML framework to automatically detect csPCa. The proposed ML framework, supported by clinical variables could be tested to assist healthcare providers in improving the performance of automated csPCa detection. Source of Funding: None © 2023 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 209Issue Supplement 4April 2023Page: e106 Advertisement Copyright & Permissions© 2023 by American Urological Association Education and Research, Inc.MetricsAuthor Information Andre Luis Abreu More articles by this author Giovanni Cacciamani More articles by this author Masatomo Kaneko More articles by this author Vasileios Magouliantis More articles by this author Yijing Yang More articles by this author Vinay Duddalwar More articles by this author C-C Jay Kuo More articles by this author Inderbir Gill More articles by this author Chrysostomos L. Nikias More articles by this author Expand All Advertisement PDF downloadLoading ...

A New RF Sensing Framework for Human Detection Through the Wall

A recent trend for human activity recognition, fall detection, health-care applications, etc., which is using the non-contact radio-frequency (RF) signals, is known as the RF sensing. State-of-the-art radar and Wi-Fi based human sensing systems utilize high bandwidth and require high computational burden. In this paper, a machine learning (ML) framework is proposed for device-free through-the-wall (TTW) human detection using narrow-band continuous RF signals that are transmitted and captured by low-cost software-defined radio (SDR) modules. To be more realistic the double TTW scenario, where both transmitter and receiver are placed outside the walls of the monitored area, is considered. Three different classes are defined as empty (human-free), moving human, and additionally the static human (only breathing, not moving) which addresses the unconscious or sleeping person. In this framework, the variance of the respiratory rate (RR) estimates of the subject and overall signal variances are used as distinguishing features. RR estimates are obtained using nonlinear least-squares (NLS) approach. It is shown in the experiments conducted with real measurements from different environments that the proposed ML framework achieves over 99% classification accuracy using a Decision Tree (DT) classifier. In addition, it is also shown that the proposed framework achieves over 95% accuracy in a completely different environment where the model is not trained. This indicates that the proposed ML framework can be applied generally without requiring specific re-training for different environments and conditions.