J Mol Model. 2024 Feb 10;30(3):65. doi: 10.1007/s00894-024-05834-2.

ABSTRACT

CONTEXT: With the union of machine learning (ML) and quantum chemistry, amid the debate between machine-learned functionals and human-designed functionals in density functional theory (DFT), this paper aims to demonstrate the generation of potential energy surfaces using computations with machine-learned density functional approximation (ML-DFA). A recent research trend is the application of ML in quantum sciences in the design of density functionals such as DeepMind's Deep Learning model (DeepMind21, DM21). Though science reported the state-of-the-art performance of DM21, the opportunity to utilize DeepMind's pretrained DM21 neural networks in computations in quantum chemistry has not yet been tapped. So far in the literature, the Deep Learning density functionals (DM21) have not been applied to generate potential energy surfaces. While the superior accuracy of DM21 has been reported, there is still a scarcity of publications that apply DM21 in calculations in the field. In this context, for the first time in literature, neural density functionals inferring 2D potential energy surfaces (ML-DFA-PES) based on machine-learned DFA-based computational method is contributed in this paper. This paper reports the ML-DFA-generated PES for C4H8, H2O, H2, and H2+ by employing a pretrained DM21m TensorFlow model with cc-pVDZ basis set. In addition, we also analyze the long-range behavior of DM21 based PES to investigate the ability to describe a system at long ranges. Furthermore, we compare PES diagrams from DM21 with popular DFT functionals (b3lyp/ PW6B95) and CCSD(T).

METHODS: In this method, 2D potential energy surfaces are obtained using a method that relies upon the neural network's ability to accurately learn the mapping between 3D electron density and exchange-correlation potential. By inserting Deep Learning inference in DFT with a pretrained neural network, self-consistent field (SCF) energy at different geometries along the coordinates of interest is computed, and then, potential energy surfaces are plotted. In this method, first, the electron density is computed mathematically, and this computed 3D electron density is used as a ML feature vector to predict the exchange correlation potential as a ML inference computed by a forward pass of pre-trained DM21 TensorFlow computational graph, followed by the computation of self-consistent field energy at multiple geometries, and then, SCF energies at different bond lengths/angles are plotted as 2D PES. We implement this in a python source code using frameworks such as PySCF and DM21. This paper contributes this implementation in open source. The source code and DM21-DFA-based PES are contributed at https://sites.google.com/view/MLfunctionals-DeepMind-PES .

PMID:38340208 | DOI:10.1007/s00894-024-05834-2

JACC Clin Electrophysiol. 2024 Jan 25:S2405-500X(24)00019-7. doi: 10.1016/j.jacep.2024.01.008. Online ahead of print.

ABSTRACT

BACKGROUND: Continuous monitoring for atrial fibrillation (AF) using photoplethysmography (PPG) from smartwatches or other wearables is challenging due to periods of poor signal quality during motion or suboptimal wearing. As a result, many consumer wearables sample infrequently and only analyze when the user is at rest, which limits the ability to perform continuous monitoring or to quantify AF.

OBJECTIVES: This study aimed to compare 2 methods of continuous monitoring for AF in free-living patients: a well-validated signal processing (SP) heuristic and a convolutional deep neural network (DNN) trained on raw signal.

METHODS: We collected 4 weeks of continuous PPG and electrocardiography signals in 204 free-living patients. Both SP and DNN models were developed and validated both on holdout patients and an external validation set.

RESULTS: The results show that the SP model demonstrated receiver-operating characteristic area under the curve (AUC) of 0.972 (sensitivity 99.6%, specificity: 94.4%), which was similar to the DNN receiver-operating characteristic AUC of 0.973 (sensitivity 92.2, specificity: 95.5%); however, the DNN classified significantly more data (95% vs 62%), revealing its superior tolerance of tracings prone to motion artifact. Explainability analysis revealed that the DNN automatically suppresses motion artifacts, evaluates irregularity, and learns natural AF interbeat variability. The DNN performed better and analyzed more signal in the external validation cohort using a different population and PPG sensor (AUC, 0.994; 97% analyzed vs AUC, 0.989; 88% analyzed).

CONCLUSIONS: DNNs perform at least as well as SP models, classify more data, and thus may be better for continuous PPG monitoring.

PMID:38340117 | DOI:10.1016/j.jacep.2024.01.008

Comput Methods Biomech Biomed Engin. 2024 Feb 10:1-23. doi: 10.1080/10255842.2024.2313066. Online ahead of print.

ABSTRACT

Multimodal sentiment analysis, an increasingly vital task in the realms of natural language processing and machine learning, addresses the nuanced understanding of emotions and sentiments expressed across diverse data sources. This study presents the Hybrid LXGB (Long short-term memory Extreme Gradient Boosting) Model, a novel approach for multimodal sentiment analysis that merges the strengths of long short-term memory (LSTM) and XGBoost classifiers. The primary objective is to address the intricate task of understanding emotions across diverse data sources, such as textual data, images, and audio cues. By leveraging the capabilities of deep learning and gradient boosting, the Hybrid LXGB Model achieves an exceptional accuracy of 97.18% on the CMU-MOSEI dataset, surpassing alternative classifiers, including LSTM, CNN, DNN, and XGBoost. This study not only introduces an innovative model but also contributes to the field by showcasing its effectiveness and balance in capturing the nuanced spectrum of sentiments within multimodal datasets. The comparison with equivalent studies highlights the model's remarkable success, emphasizing its potential for practical applications in real-world scenarios. The Hybrid LXGB Model offers a unique and promising perspective in the realm of multimodal sentiment analysis, demonstrating the significance of integrating LSTM and XGBoost for enhanced performance.

PMID:38340022 | DOI:10.1080/10255842.2024.2313066

Expert Rev Clin Pharmacol. 2024 Feb 10. doi: 10.1080/17512433.2024.2317963. Online ahead of print.

ABSTRACT

INTRODUCTION: The integration of artificial intelligence (AI) into clinical pharmacology could be a potential approach for accelerating drug discovery and development, improving patient care, and streamlining medical research processes.

AREAS COVERED: We reviewed the current state of AI applications in clinical pharmacology, focusing on drug discovery and development, precision medicine, pharmacovigilance, and other ventures. Key AI applications in clinical pharmacology are examined, including machine learning, natural language processing, deep learning, and reinforcement learning etc. Additionally, the evolving role of clinical pharmacologists, ethical considerations, and challenges in implementing AI in clinical pharmacology are discussed.

EXPERT OPINION: The AI could be instrumental in accelerating drug discovery, predicting drug safety and efficacy, and optimizing clinical trial designs. It can play a vital role in precision medicine by helping in personalized drug dosing, treatment selection, and predicting drug response based on genetic, clinical, and environmental factors. The role of AI in pharmacovigilance, such as signal detection and adverse event prediction, is also promising. The collaboration between clinical pharmacologists and AI experts also poses certain ethical and practical challenges. Clinical pharmacologists can be instrumental in shaping the future of AI-driven clinical pharmacology and contribute to the improvement of healthcare systems.

PMID:38340012 | DOI:10.1080/17512433.2024.2317963

Technol Health Care. 2024 Feb 1. doi: 10.3233/THC-231807. Online ahead of print.

ABSTRACT

BACKGROUND: Coronary heart disease (CHD) is one of the deadliest diseases and a risk prediction model for cardiovascular conditions is needed. Due to the huge number of features that lead to heart problems, it is often difficult for an expert to evaluate these huge features into account. So, there is a need of appropriate feature selection for the given CHD dataset. For early CHD detection, deep learning modes (DL) show promising results in the existing studies.

OBJECTIVE: This study aimed to develop a deep convolution neural network (CNN) model for classification with a selected number of efficient features using the LASSO (least absolute shrinkage and selection operator) technique. Also, aims to compare the model with similar studies and analyze the performance of the proposed model using accuracy measures.

METHODS: The CHD dataset of NHANES (National Health and Nutritional Examination Survey) was examined with 49 features using LASSO technique. This research work is an attempt to apply an improved CNN model for the classification of the CHD dataset with huge features CNN model with feature extractor consists of a fully connected layer with two convolution 1D layers, and classifier part consists of two fully connected layers with SoftMax function was trained on this dataset. Metrics like accuracy recall, specificity, and ROC were used for the evaluation of the proposed model.

RESULTS: The feature selection was performed by applying the LASSO model. The proposed CNN model achieved 99.36% accuracy, while previous studies model achieved over 80 to 92% accuracy.

CONCLUSION: The application of the proposed CNN with the LASSO model for the classification of CHD can speed up the diagnosis of CHD and appears to be effective in predicting cardiovascular disease based on risk features.

PMID:38339946 | DOI:10.3233/THC-231807

Hum Brain Mapp. 2024 Feb 1;45(2):e26578. doi: 10.1002/hbm.26578.

ABSTRACT

Fibre tract delineation from diffusion magnetic resonance imaging (MRI) is a valuable clinical tool for neurosurgical planning and navigation, as well as in research neuroimaging pipelines. Several popular methods are used for this task, each with different strengths and weaknesses making them more or less suited to different contexts. For neurosurgical imaging, priorities include ease of use, computational efficiency, robustness to pathology and ability to generalise to new tracts of interest. Many existing methods use streamline tractography, which may require expert neuroimaging operators for setting parameters and delineating anatomical regions of interest, or suffer from as a lack of generalisability to clinical scans involving deforming tumours and other pathologies. More recently, data-driven approaches including deep-learning segmentation models and streamline clustering methods have improved reproducibility and automation, although they can require large amounts of training data and/or computationally intensive image processing at the point of application. We describe an atlas-based direct tract mapping technique called 'tractfinder', utilising tract-specific location and orientation priors. Our aim was to develop a clinically practical method avoiding streamline tractography at the point of application while utilising prior anatomical knowledge derived from only 10-20 training samples. Requiring few training samples allows emphasis to be placed on producing high quality, neuro-anatomically accurate training data, and enables rapid adaptation to new tracts of interest. Avoiding streamline tractography at the point of application reduces computational time, false positives and vulnerabilities to pathology such as tumour deformations or oedema. Carefully filtered training streamlines and track orientation distribution mapping are used to construct tract specific orientation and spatial probability atlases in standard space. Atlases are then transformed to target subject space using affine registration and compared with the subject's voxel-wise fibre orientation distribution data using a mathematical measure of distribution overlap, resulting in a map of the tract's likely spatial distribution. This work includes extensive performance evaluation and comparison with benchmark techniques, including streamline tractography and the deep-learning method TractSeg, in two publicly available healthy diffusion MRI datasets (from TractoInferno and the Human Connectome Project) in addition to a clinical dataset comprising paediatric and adult brain tumour scans. Tract segmentation results display high agreement with established techniques while requiring less than 3 min on average when applied to a new subject. Results also display higher robustness than compared methods when faced with clinical scans featuring brain tumours and resections. As well as describing and evaluating a novel proposed tract delineation technique, this work continues the discussion on the challenges surrounding the white matter segmentation task, including issues of anatomical definitions and the use of quantitative segmentation comparison metrics.

PMID:38339907 | DOI:10.1002/hbm.26578

Sensors (Basel). 2024 Feb 4;24(3):1012. doi: 10.3390/s24031012.

ABSTRACT

Optical encryption based on single-pixel imaging (SPI) has made great advances with the introduction of deep learning. However, the use of deep neural networks usually requires a long training time, and the networks need to be retrained once the target scene changes. With this in mind, we propose an SPI encryption scheme based on an attention-inserted physics-driven neural network. Here, an attention module is used to encrypt the single-pixel measurement value sequences of two images, together with a sequence of cryptographic keys, into a one-dimensional ciphertext signal to complete image encryption. Then, the encrypted signal is fed into a physics-driven neural network for high-fidelity decoding (i.e., decryption). This scheme eliminates the need for pre-training the network and gives more freedom to spatial modulation. Both simulation and experimental results have demonstrated the feasibility and eavesdropping resistance of this scheme. Thus, it will lead SPI-based optical encryption closer to intelligent deep encryption.

PMID:38339728 | DOI:10.3390/s24031012

Sensors (Basel). 2024 Feb 4;24(3):1002. doi: 10.3390/s24031002.

ABSTRACT

Identifying the classes and locations of prohibited items is the target of security inspection. However, X-ray security inspection images with insufficient feature extraction, imbalance between easy and hard samples, and occlusion lead to poor detection accuracy. To address the above problems, an object-detection method based on YOLOv8 is proposed. Firstly, an ASFF (adaptive spatial feature fusion) and a weighted feature concatenation algorithm are introduced to fully extract the scale features from input images. In this way, the model can learn further details in training. Secondly, CoordAtt (coordinate attention module), which belongs to the hybrid attention mechanism, is embedded to enhance the learning of features of interest. Then, the slide loss function is introduced to balance the simple samples and the difficult samples. Finally, Soft-NMS (non-maximum suppression) is introduced to resist the conditions containing occlusion. The experimental result shows that mAP (mean average precision) achieves 90.2%, 90.5%, 79.1%, and 91.4% on the Easy, Hard, and Hidden sets of the PIDray and SIXray public test set, respectively. Contrasted with original model, the mAP of our proposed YOLOv8n model increased by 2.7%, 3.1%, 9.3%, and 2.4%, respectively. Furthermore, the parameter count of the modified YOLOv8n model is roughly only 3 million.

PMID:38339718 | DOI:10.3390/s24031002

Sensors (Basel). 2024 Feb 3;24(3):989. doi: 10.3390/s24030989.

ABSTRACT

In recent years, significant progress has been witnessed in the field of deep learning-based object detection. As a subtask in the field of object detection, traffic sign detection has great potential for development. However, the existing object detection methods for traffic sign detection in real-world scenes are plagued by issues such as the omission of small objects and low detection accuracies. To address these issues, a traffic sign detection model named YOLOv7-Traffic Sign (YOLOv7-TS) is proposed based on sub-pixel convolution and feature fusion. Firstly, the up-sampling capability of the sub-pixel convolution integrating channel dimension is harnessed and a Feature Map Extraction Module (FMEM) is devised to mitigate the channel information loss. Furthermore, a Multi-feature Interactive Fusion Network (MIFNet) is constructed to facilitate enhanced information interaction among all feature layers, improving the feature fusion effectiveness and strengthening the perception ability of small objects. Moreover, a Deep Feature Enhancement Module (DFEM) is established to accelerate the pooling process while enriching the highest-layer feature. YOLOv7-TS is evaluated on two traffic sign datasets, namely CCTSDB2021 and TT100K. Compared with YOLOv7, YOLOv7-TS, with a smaller number of parameters, achieves a significant enhancement of 3.63% and 2.68% in the mean Average Precision (mAP) for each respective dataset, proving the effectiveness of the proposed model.

PMID:38339706 | DOI:10.3390/s24030989

Sensors (Basel). 2024 Feb 2;24(3):987. doi: 10.3390/s24030987.

ABSTRACT

This paper introduces an approach to the automated measurement and analysis of dairy cows using 3D point cloud technology. The integration of advanced sensing techniques enables the collection of non-intrusive, precise data, facilitating comprehensive monitoring of key parameters related to the health, well-being, and productivity of dairy cows. The proposed system employs 3D imaging sensors to capture detailed information about various parts of dairy cows, generating accurate, high-resolution point clouds. A robust automated algorithm has been developed to process these point clouds and extract relevant metrics such as dairy cow stature height, rump width, rump angle, and front teat length. Based on the measured data combined with expert assessments of dairy cows, the quality indices of dairy cows are automatically evaluated and extracted. By leveraging this technology, dairy farmers can gain real-time insights into the health status of individual cows and the overall herd. Additionally, the automated analysis facilitates efficient management practices and optimizes feeding strategies and resource allocation. The results of field trials and validation studies demonstrate the effectiveness and reliability of the automated 3D point cloud approach in dairy farm environments. The errors between manually measured values of dairy cow height, rump angle, and front teat length, and those calculated by the auto-measurement algorithm were within 0.7 cm, with no observed exceedance of errors in comparison to manual measurements. This research contributes to the burgeoning field of precision livestock farming, offering a technological solution that not only enhances productivity but also aligns with contemporary standards for sustainable and ethical animal husbandry practices.

PMID:38339704 | DOI:10.3390/s24030987

Sensors (Basel). 2024 Feb 2;24(3):984. doi: 10.3390/s24030984.

ABSTRACT

In the process of industrial production, manual assembly of workpieces exists with low efficiency and high intensity, and some of the assembly process of the human body has a certain degree of danger. At the same time, traditional machine learning algorithms are difficult to adapt to the complexity of the current industrial field environment; the change in the environment will greatly affect the accuracy of the robot's work. Therefore, this paper proposes a method based on the combination of machine vision and the YOLOv5 deep learning model to obtain the disk porous localization information, after coordinate mapping by the ROS communication control robotic arm work, in order to improve the anti-interference ability of the environment and work efficiency but also reduce the danger to the human body. The system utilizes a camera to collect real-time images of targets in complex environments and, then, trains and processes them for recognition such that coordinate localization information can be obtained. This information is converted into coordinates under the robot coordinate system through hand-eye calibration, and the robot is then controlled to complete multi-hole localization and tracking by means of communication between the upper and lower computers. The results show that there is a high accuracy in the training and testing of the target object, and the control accuracy of the robotic arm is also relatively high. The method has strong anti-interference to the complex environment of industry and exhibits a certain feasibility and effectiveness. It lays a foundation for achieving the automated installation of docking disk workpieces in industrial production and also provides a more favorable choice for the production and installation of the process of screw positioning needs.

PMID:38339701 | DOI:10.3390/s24030984

Sensors (Basel). 2024 Feb 2;24(3):980. doi: 10.3390/s24030980.

ABSTRACT

The development of emerging information technologies, such as the Internet of Things (IoT), edge computing, and blockchain, has triggered a significant increase in IoT application services and data volume. Ensuring satisfactory service quality for diverse IoT application services based on limited network resources has become an urgent issue. Generalized processor sharing (GPS), functioning as a central resource scheduling mechanism guiding differentiated services, stands as a key technology for implementing on-demand resource allocation. The performance prediction of GPS is a crucial step that aims to capture the actual allocated resources using various queue metrics. Some methods (mainly analytical methods) have attempted to establish upper and lower bounds or approximate solutions. Recently, artificial intelligence (AI) methods, such as deep learning, have been designed to assess performance under self-similar traffic. However, the proposed methods in the literature have been developed for specific traffic scenarios with predefined constraints, thus limiting their real-world applicability. Furthermore, the absence of a benchmark in the literature leads to an unfair performance prediction comparison. To address the drawbacks in the literature, an AI-enabled performance benchmark with comprehensive traffic-oriented experiments showcasing the performance of existing methods is presented. Specifically, three types of methods are employed: traditional approximate analytical methods, traditional machine learning-based methods, and deep learning-based methods. Following that, various traffic flows with different settings are collected, and intricate experimental analyses at both the feature and method levels under different traffic conditions are conducted. Finally, insights from the experimental analysis that may be beneficial for the future performance prediction of GPS are derived.

PMID:38339696 | DOI:10.3390/s24030980

Sensors (Basel). 2024 Feb 2;24(3):970. doi: 10.3390/s24030970.

ABSTRACT

In this paper, we present the development of a low-cost distributed computing pipeline for cotton plant phenotyping using Raspberry Pi, Hadoop, and deep learning. Specifically, we use a cluster of several Raspberry Pis in a primary-replica distributed architecture using the Apache Hadoop ecosystem and a pre-trained Tiny-YOLOv4 model for cotton bloom detection from our past work. We feed cotton image data collected from a research field in Tifton, GA, into our cluster's distributed file system for robust file access and distributed, parallel processing. We then submit job requests to our cluster from our client to process cotton image data in a distributed and parallel fashion, from pre-processing to bloom detection and spatio-temporal map creation. Additionally, we present a comparison of our four-node cluster performance with centralized, one-, two-, and three-node clusters. This work is the first to develop a distributed computing pipeline for high-throughput cotton phenotyping in field-based agriculture.

PMID:38339687 | DOI:10.3390/s24030970

Sensors (Basel). 2024 Feb 1;24(3):958. doi: 10.3390/s24030958.

ABSTRACT

The integration of artificial intelligence (AI) with Digital Twins (DTs) has emerged as a promising approach to revolutionize healthcare, particularly in terms of diagnosis and management of thoracic disorders. This study proposes a comprehensive framework, named Lung-DT, which leverages IoT sensors and AI algorithms to establish the digital representation of a patient's respiratory health. Using the YOLOv8 neural network, the Lung-DT system accurately classifies chest X-rays into five distinct categories of lung diseases, including "normal", "covid", "lung_opacity", "pneumonia", and "tuberculosis". The performance of the system was evaluated employing a chest X-ray dataset available in the literature, demonstrating average accuracy of 96.8%, precision of 92%, recall of 97%, and F1-score of 94%. The proposed Lung-DT framework offers several advantages over conventional diagnostic methods. Firstly, it enables real-time monitoring of lung health through continuous data acquisition from IoT sensors, facilitating early diagnosis and intervention. Secondly, the AI-powered classification module provides automated and objective assessments of chest X-rays, reducing dependence on subjective human interpretation. Thirdly, the twin digital representation of the patient's respiratory health allows for comprehensive analysis and correlation of multiple data streams, providing valuable insights as to personalized treatment plans. The integration of IoT sensors, AI algorithms, and DT technology within the Lung-DT system demonstrates a significant step towards improving thoracic healthcare. By enabling continuous monitoring, automated diagnosis, and comprehensive data analysis, the Lung-DT framework has enormous potential to enhance patient outcomes, reduce healthcare costs, and optimize resource allocation.

PMID:38339678 | DOI:10.3390/s24030958

Biomed Phys Eng Express. 2024 Feb 9. doi: 10.1088/2057-1976/ad27ea. Online ahead of print.

ABSTRACT

Macular Edema is a leading cause of visual impairment and blindness in patients with ocular fundus diseases. Due to its non-invasive and high-resolution characteristics, optical coherence tomography (OCT) has been extensively utilized for the diagnosis of macular diseases. The manual detection of retinal diseases by clinicians is a laborious process, further complicated by the challenging identification of macular diseases. This difficulty arises from the significant pathological alterations occurring within the retinal layers, as well as the accumulation of fluid in the retina. Deep Learning neural networks are utilized for automatic detection of retinal diseases. This paper aims to propose a lightweight hybrid learning Retinal Disease OCT Net with a reduced number of trainable parameters and enable automatic classification of retinal diseases. A Hybrid Learning Retinal Disease OCT Net (RD-OCT) is utilized for the multiclass classification of major retinal diseases, namely neovascular age-related macular degeneration (nAMD), diabetic macular edema (DME), retinal vein occlusion (RVO), and normal retinal conditions. The diagnosis of retinal diseases is facilitated by the use of hybrid learning models and pre-trained deep learning models in the field of artificial intelligence. The Hybrid Learning RD-OCT Net provides better accuracy of 97.6% for nAMD, 98.08% for DME, 98% for RVO, and 97% for the Normal group. The respective area under the curve values were 0.99, 0.97, 1.0, and 0.99. The utilization of the RD-OCT model will be useful for ophthalmologists in the diagnosis of prevalent retinal diseases, due to the simplicity of the system and reduced number of trainable parameters.

PMID:38335542 | DOI:10.1088/2057-1976/ad27ea

PLoS One. 2024 Feb 9;19(2):e0296979. doi: 10.1371/journal.pone.0296979. eCollection 2024.

ABSTRACT

With the rapid development of smart grids, society has become increasingly urgent to solve the problems of low energy utilization efficiency and high energy consumption. In this context, load identification has become a key element in formulating scientific and effective energy consumption plans and reducing unnecessary energy waste. However, traditional load identification methods mainly focus on known electrical equipment, and accurate identification of unknown electrical equipment still faces significant challenges. A new encoding feature space based on Triplet neural networks is proposed in this paper to detect unknown electrical appliances using convex hull coincidence degree. Additionally, transfer learning is introduced for the rapid updating of the pre-classification model's self-incrementing class with the unknown load. In experiments, the effectiveness of our method is successfully tested on the PLAID dataset. The accuracy of unknown load identification reached 99.23%. Through this research, we expect to bring a new idea to the field of load identification to meet the urgent need for the identification of unknown electrical appliances in the development of smart grids.

PMID:38335185 | DOI:10.1371/journal.pone.0296979

Med Phys. 2024 Feb 9. doi: 10.1002/mp.16964. Online ahead of print.

ABSTRACT

BACKGROUND: Notwithstanding the encouraging results of previous studies reporting on the efficiency of deep learning (DL) in COVID-19 prognostication, clinical adoption of the developed methodology still needs to be improved. To overcome this limitation, we set out to predict the prognosis of a large multi-institutional cohort of patients with COVID-19 using a DL-based model.

PURPOSE: This study aimed to evaluate the performance of deep privacy-preserving federated learning (DPFL) in predicting COVID-19 outcomes using chest CT images.

METHODS: After applying inclusion and exclusion criteria, 3055 patients from 19 centers, including 1599 alive and 1456 deceased, were enrolled in this study. Data from all centers were split (randomly with stratification respective to each center and class) into a training/validation set (70%/10%) and a hold-out test set (20%). For the DL model, feature extraction was performed on 2D slices, and averaging was performed at the final layer to construct a 3D model for each scan. The DensNet model was used for feature extraction. The model was developed using centralized and FL approaches. For FL, we employed DPFL approaches. Membership inference attack was also evaluated in the FL strategy. For model evaluation, different metrics were reported in the hold-out test sets. In addition, models trained in two scenarios, centralized and FL, were compared using the DeLong test for statistical differences.

RESULTS: The centralized model achieved an accuracy of 0.76, while the DPFL model had an accuracy of 0.75. Both the centralized and DPFL models achieved a specificity of 0.77. The centralized model achieved a sensitivity of 0.74, while the DPFL model had a sensitivity of 0.73. A mean AUC of 0.82 and 0.81 with 95% confidence intervals of (95% CI: 0.79-0.85) and (95% CI: 0.77-0.84) were achieved by the centralized model and the DPFL model, respectively. The DeLong test did not prove statistically significant differences between the two models (p-value = 0.98). The AUC values for the inference attacks fluctuate between 0.49 and 0.51, with an average of 0.50 ± 0.003 and 95% CI for the mean AUC of 0.500 to 0.501.

CONCLUSION: The performance of the proposed model was comparable to centralized models while operating on large and heterogeneous multi-institutional datasets. In addition, the model was resistant to inference attacks, ensuring the privacy of shared data during the training process.

PMID:38335175 | DOI:10.1002/mp.16964

IEEE Trans Vis Comput Graph. 2024 Feb 9;PP. doi: 10.1109/TVCG.2024.3364365. Online ahead of print.

ABSTRACT

This study presents a self-prior-based mesh inpainting framework that requires only an incomplete mesh as input, without the need for any training datasets. Additionally, our method maintains the polygonal mesh format throughout the inpainting process without converting the shape format to an intermediate, such as a voxel grid, a point cloud, or an implicit function, which are typically considered easier for deep neural networks to process. To achieve this goal, we introduce two graph convolutional networks (GCNs): single-resolution GCN (SGCN) and multi-resolution GCN (MGCN), both trained in a self-supervised manner. Our approach refines a watertight mesh obtained from the initial hole filling to generate a completed output mesh. Specifically, we train the GCNs to deform an oversmoothed version of the input mesh into the expected completed shape. To supervise the GCNs for accurate vertex displacements, despite the unknown correct displacements at real holes, we utilize multiple sets of meshes with several connected regions marked as fake holes. The correct displacements are known for vertices in these fake holes, enabling network training with loss functions that assess the accuracy of displacement vectors estimated by the GCNs. We demonstrate that our method outperforms traditional dataset-independent approaches and exhibits greater robustness compared to other deep-learning-based methods for shapes that less frequently appear in shape datasets.

PMID:38335080 | DOI:10.1109/TVCG.2024.3364365

IEEE Trans Med Imaging. 2024 Feb 9;PP. doi: 10.1109/TMI.2024.3364911. Online ahead of print.

ABSTRACT

Magnetic resonance imaging is subject to slow acquisition times due to the inherent limitations in data sampling. Recently, supervised deep learning has emerged as a promising technique for reconstructing sub-sampled MRI. However, supervised deep learning requires a large dataset of fully-sampled data. Although unsupervised or self-supervised deep learning methods have emerged to address the limitations of supervised deep learning approaches, they still require a database of images. In contrast, scan-specific deep learning methods learn and reconstruct using only the sub-sampled data from a single scan. Here, we introduce Scan-Specific Self-Supervised Bayesian Deep Non-Linear Inversion (DNLINV) that does not require an auto calibration scan region. DNLINV utilizes a Deep Image Prior-type generative modeling approach and relies on approximate Bayesian inference to regularize the deep convolutional neural network. We demonstrate our approach on several anatomies, contrasts, and sampling patterns and show improved performance over existing approaches in scan-specific calibrationless parallel imaging and compressed sensing.

PMID:38335079 | DOI:10.1109/TMI.2024.3364911

IEEE Trans Biomed Circuits Syst. 2024 Feb 9;PP. doi: 10.1109/TBCAS.2024.3364235. Online ahead of print.

ABSTRACT

Deep learning (DL) has been used for electromyographic (EMG) signal recognition and achieved high accuracy for multiple classification tasks. However, implementation in resource-constrained prostheses and human-computer interaction devices remains challenging. To overcome these problems, this paper implemented a low-power system for EMG gesture and force level recognition using Zynq architecture. Firstly, a lightweight network model structure was proposed by Ultra-lightweight depth separable convolution (UL-DSC) and channel attention-global average pooling (CA-GAP) to reduce the computational complexity while maintaining accuracy. A wearable EMG acquisition device for real-time data acquisition was subsequently developed with size of 36mm×28mm×4mm. Finally, a highly parallelized dedicated hardware accelerator architecture was designed for inference computation. 18 gestures were tested, including force levels from 22 healthy subjects. The results indicate that the average accuracy rate was 94.92% for a model with 5.0k parameters and a size of 0.026MB. Specifically, the average recognition accuracy for static and force-level gestures was 98.47% and 89.92%, respectively. The proposed hardware accelerator architecture was deployed with 8-bit precision, a single-frame signal inference time of 41.9μs, a power consumption of 0.317W, and a data throughput of 78.6 GOP/s.

PMID:38335070 | DOI:10.1109/TBCAS.2024.3364235