Neural Network Architecture Research Articles

Introduction: Myocardial infarct size (IS) is the most robust endpoint for evaluating cardioprotective strategies in preclinical ischemia/reperfusion studies. The gold standard for IS quantification in preclinical studies (triphenyl tetrazolium chloride (TTC) staining) is traditionally performed manually and is prone to inter-operator variability. Here, we propose a deep learning segmentation pipeline to automate IS quantification in TTC-stained rat heart sections. Methods: We used n=165 Sprague-Dawley rats (150–300 g, 1–2 months, 69% female). Myocardial infarction (MI) was induced using a standard occlusion/reperfusion model by occluding the proximal left coronary artery for 30 minutes, followed by 3 hours of reperfusion. After euthanasia, the left ventricle (LV) was excised, transversely sliced, and incubated in 1% TTC at 37 °C for 15 minutes to distinguish necrotic myocardium (pale white) from viable tissues (brick red, Fig. 1). Manual IS was quantified by contouring infarcted and total LV areas in each slice using ImageJ (NIH, USA). To automate the IS measurement from TTC-stained heart slices, we implemented a deep learning segmentation pipeline based on the mask region-based convolutional neural network (Mask R-CNN) architecture. Ground truth masks for infarcted regions and LV area were created using VGG Image Annotator. Images from n=140 rats were used for training, as well as an additional 1,400 images generated by data augmentation. All training and preprocessing pipelines were conducted in Python. Dice similarity coefficient (Dice score) was used to evaluate the model performance. The best-performing Mask R-CNN model was blindly tested on 25 additional MI rats. Results: Infarct sizes calculated from Mask R-CNN-generated segmentations showed strong agreement with the ones from expert-annotated manual segmentations from TTC-stained LV slices (R = 0.97, p < 0.0001) when tested on heart slices from 25 additional MI rats, supporting the model’s accuracy and validity. Conclusions: Our results demonstrate that deep learning segmentation accurately and automatically quantifies infarct size from TTC-stained images without operator input. This automated approach is rapid, reproducible, and unbiased, significantly reducing inter-operator variability and manual workload in preclinical studies. By streamlining infarct size assessment in preclinical cardio-protection studies, it has the potential to improve consistency and translational value in cardiac research.

Background: Left ventricular filling pressure (LVFP) is associated with symptoms and signs of heart failure, can guide therapeutic decision-making, and provides prognostic insights. Purpose: We aimed to develop an artificial intelligence (AI)-based electrocardiogram (ECG) model to detect increased LVFP and evaluate its prognostic significance through multi-ethnic geographically diverse cohorts. Methods: The septal E/e' value over 15 on Doppler echocardiography was used to define increased LVFP and to guide model training. The AI-ECG model, based on a transformer-enhanced convolutional neural network architecture, was initially pre-trained using a multinational public ECG dataset and subsequently fine-tuned as a binary classifier using a development cohort comprising 225,737 12-lead ECGs and 115,982 echocardiograms from 92,775 unique patients across two tertiary hospitals in Korea. The model performance to discriminate increased LVFP and left ventricular diastolic dysfunction grade (LVDD) II/III were assessed in an internal test cohort (n=9,278) and an independent external test cohort (n=17,926) from a third tertiary Korean hospital. The prognostic significance of the AI-ECG model was validated via survival analyses in these two cohorts and the prognostic significance validation cohort (UK biobank, n=43,347). Results: The AI-ECG model detected increased LVFP with area under the receiver operating characteristic curves (AUROCs) of 0.868 (95% confidence interval, 0.859–0.877) and 0.850 (0.841–0.858) in the internal and external test cohort, respectively. For detecting LVDD grade II/III, the model demonstrated an AUROCs of 0.901 (0.892–0.910) and 0.941 (0.931–0.951) in the internal and external test cohort, respectively. The model’s prognostic utility was confirmed across all three cohorts by Kaplan-Meier survival analyses, demonstrating significantly higher mortality in patients with higher AI-ECG scores. In Cox models adjusted for age, sex, and clinical comorbidities, the AI-ECG score remained independently associated with mortality across all cohorts: adjusted hazard ratio 1.31 (95% CI, 1.23–1.38, p<0.001) in the internal test cohort, 1.32 (1.28–1.35, p<0.001) in the external test cohort, and 1.16 (1.07–1.26, p<0.001) in the UK biobank. Conclusion: We developed a multiethnic foundation model-based, generalizable AI-ECG capable of detecting increased LVFP and prognosticating long-term mortality, with external validation across multiethnic cohorts.

Background: Natriuretic peptides such as BNP and NT-proBNP remain cornerstone biomarkers for diagnosing and management of heart failure (HF). However, their measurement is done mainly when HF is suspected or requires monitoring, posing a missed opportunity for early HF detection and proactive intervention in at-risk patients. This gap highlights the need for accessible, non-invasive tools that can support earlier screening and continuous assessment. Aim: To develop and validate multi-task deep learning models that simultaneously estimate natriuretic peptide values and classify clinically relevant strata directly from raw 12-lead and Lead I ECG waveforms. Methods: Two distinct one-dimensional residual convolutional neural network architectures were separately trained on 12-lead and Lead I ECG inputs, using paired ECG and BNP measurements within a ±2-hour window to ensure rigorous label alignment. The models performed both BNP regression and classification into clinically meaningful low (<100 pg/mL) and elevated (>500 pg/mL) strata. Internal validation used a holdout cohort from Wake Forest Baptist Health (WF) (Winston-Salem, NC), while external validation employed an independent cohort from the University of Tennessee Health Science Center (UTHSC) (Memphis, TN), which included NT-proBNP data and specific clinical thresholds (low <125 pg/mL and elevated >300 pg/mL). Performance was evaluated using AUC, PPV, and NPV for classification and Spearman correlation for regression. Results: The internal dataset (WF) included 102,311 paired ECG–BNP samples from 54,526 patients, with a holdout set of 10,264 samples for validation. The external cohort (UTHSC) comprised 88,179 same day ECG–NT-proBNP pairs. The multi-task models demonstrated strong classification accuracy and regression correlation across both datasets. Specifically, the 12-lead model achieved AUCs of 0.88–0.89 and Spearman correlations of 0.75 in internal BNP strata, with similar performance in the external NT-proBNP cohort (AUC 0.88, Spearman 0.75–0.76). The Lead I model showed slightly lower but robust performance. Detailed metrics with 95% confidence intervals are summarized in Table 1. Conclusion: This multi-task models provide accurate, simultaneous estimation and classification of natriuretic peptides from ECGs. This approach offers a rapid, non-invasive tool for HF biomarker assessment that may improve clinical workflows, enable ambulatory monitoring, and enhance timely clinical decision-making.

Introduction: Severe aortic regurgitation (AR) is characterized by significant retrograde blood flow in the aorta and remains difficult to quantitively evaluate by echocardiography. By providing comprehensive insights into hemodynamic changes and quantifying regurgitant fraction (RF) across various locations of the aorta, this study investigated the potential of 4D flow MRI to enhance diagnostic accuracy and inform clinical decision-making. Methods: An institutional database was queried for patients with chronic AR on echocardiography and paired cardiac MRIs with aortic 4D flow MRI. Patients with LVEF < 50%, concomitant mitral regurgitation and aortic stenosis were excluded. A fully automated 4D flow MRI processing tool, performing standard preprocessing corrections and aortic 3D segmentation using separately trained machine learning models (Dense U-net convolutional neural network architecture) was used. Through-plane flow was quantified at 7 AHA-standardized locations: aortic annulus, sinotubular junction, mid ascending aorta, distal ascending aorta, aortic arch, proximal descending aorta and mid descending aorta. 4D flow MRI-based quantifications of RF were assessed for differentiating severe AR, using echo gradings as reference classification. Adjudicated clinical outcome data included cardiac-related hospitalizations such as heart failure, arrhythmias, and inpatient management of valve intervention. Results: Of 59 patients with chronic AR, the mean age was 49 ± 14.5 years, LVEF 56.5 ± 8.3%, LV end diastolic volume 251 ± 74 mL, 90% male and 73% had bicuspid aortic valves. Receiver operator characteristic (ROC) analysis of 4D flow MRI RFs revealed the optimal anatomic location to differentiate severe AR, as graded by echo was the mid descending aorta (AUC = 0.79). In patients with moderate, moderate-severe, and severe AR on echo, Kaplan-Meyer analysis reveals significant differences in cardiac-related hospitalization rates and time to valve intervention when patients were median split by optimal mid-descending aorta ROC RF (35%) but not at other locations of the aorta nor RFs calculated by traditional 2D Phase Contrast MRI (Figure 1). Conclusion: The optimal location in discerning severe aortic regurgitation as per RF by 4D flow analysis is the mid-descending aorta. 4D flow quantified RF of 35% at the mid-descending aorta was associated with cardiac related hospitalizations.

The subject of study in this article is neural network–based methods for aerial image matching, which are widely used in navigation, localization, and mapping tasks. A key challenge lies in the sensitivity of such methods to visual disturbances and scene novelty caused by shadows, illumination changes, and terrain variability, which limits their robustness in real-world conditions—particularly under constrained computational resources. This paper investigates an approach to enhancing the robustness and cross-domain generalization of computationally efficient aerial image matching models by combining adversarial procedural noise with a modified activation function. The goal is to develop a training methodology that simultaneously increases the resilience of models to perturbations and improves their transferability across different observation domains. The research objectives are as follows: (1) to analyze existing methods for improving the robustness of neural networks and assess their applicability to aerial image matching tasks; (2) to develop a training approach incorporating the synthesis of adversarial procedural noises (Perlin, Gabor, Worley) and the replacement of the standard ReLU with a hybrid activation function, LeakyReLU6, which constrains activation amplitudes and reduces sensitivity to local disturbances; (3) to conduct a comprehensive experimental evaluation of detector-based architectures (SuperPoint + LightGlue) and detector-free models (EfficientLoFTR) using the Aerial Image Matching Benchmark dataset; (4) to verify cross-domain generalization on the HPatches dataset; and (5) to perform an ablation study to isolate the contribution of each component. Results. The proposed methodology achieved over a 4.2% absolute improvement in AUC@1px matching accuracy on noisy test data for both classes of models. The ablation study revealed a synergistic effect from combining procedural noise with LeakyReLU6 — in particular, for the SuperPoint + LightGlue combination, improvements reached +3.0% AUC@1px and +2.7% AUC@3px, while for EfficientLoFTR, gains of +2.2% and +2.6% were observed, respectively. Additionally, testing on HPatches showed a 0.83% smaller performance drop compared to baseline training, confirming a higher level of cross-domain generalization. Conclusions. The proposed approach enhances the noise robustness and cross-domain generalization of feature-matching models and can be easily extended to various neural network architectures. Future work will focus on investigating the influence of procedural noise hyperparameters, applying meta-learning on corrupted data, and introducing architectural improvements to further strengthen resilience and robustness. Scientific novelty. The novelty of this work lies in the first integration of adversarial learning with procedural noise and a bounded activation function (LeakyReLU6, using the Straight-Through Estimator (STE) in the backward pass), which produced a synergistic effect that improved the robustness and generalization of aerial image matching models without a significant increase in computational cost.

Concerning shrimp diseases, including acute hepatopancreatic necrosis disease (AHPND), hepatopancreatic parvovirus (HPV) infection and Enterocytozoon hepatopenaei (EHP) microsporidiosis negatively impact shrimp aquaculture through acute mortality, chronic growth retardation or compromised health that increases susceptibility to concurrent infections. All three diseases damage hepatopancreas, a vital organ for nutrient absorption and growth, though their clinical outcomes differ: AHPND is typically associated with rapid, high mortality, EHP primarily causes chronic production losses and HPV, while currently of lower pathogenic significance, may still impair health under certain conditions. Outbreak severity is often intensified by poor water quality, inadequate farm management, antibiotic misuse and pathogen vectors, leading to substantial economic losses. Timely and accurate diagnosis is therefore critical for effective disease management. This study investigates two convolutional neural network (CNN) architectures, EfficientNet and MobileNet. A curated and preprocessed dataset was used to fine-tune both models with a standardized custom classification head, ensuring a controlled backbone comparison. Experimental results show both architectures achieving over 95% accuracy, with MobileNet providing faster inference suitable for on-site deployment. These findings demonstrate the practical feasibility of lightweight CNN-based diagnostics tools for real-time, scalable, and cost-efficient health monitoring in shrimp aquaculture, bridging the gap between the laboratory-grade performance and field-level usability.

The detection of hate speech on social media has become a pressing challenge, particularly in multilingual and low-resource language settings such as Roman Urdu, where informal grammar, code-switching, and inconsistent orthography hinder accurate classification. Despite progress in hate speech detection for high-resource languages, limited research exists for Roman Urdu content. This study addresses this gap by proposing a computationally efficient deep learning framework based on a hybrid convolutional neural network and bidirectional long short-term memory (CNN-BiLSTM) architecture. The model leverages FastText pre-trained embeddings to capture subword-level semantics and combines convolutional layers for local feature extraction with BiLSTM for global context modeling. We evaluate our approach on a labeled Roman Urdu dataset and compare it with traditional machine learning models and deep learning baselines. Our proposed CNN-BiLSTM model achieves the highest performance with an accuracy of 80.67% and an F1-score of 81.47%, outperforming competitive baselines. These findings demonstrate the effectiveness and practicality of our lightweight architecture in detecting hate speech in Roman Urdu, offering a scalable solution for multilingual and resource-constrained environments.

The Cellular-Potts model is a powerful and ubiquitous framework for developing computational models for simulating complex multicellular biological systems. Cellular-Potts models (CPMs) are often computationally expensive due to the explicit modeling of interactions among large numbers of individual model agents and diffusive fields described by partial differential equations (PDEs). In this work, we develop a convolutional neural network (CNN) surrogate model using a U-Net architecture that accounts for periodic boundary conditions. We use this model to accelerate the evaluation of a mechanistic CPM previously used to investigate in vitro vasculogenesis. The surrogate model was trained to predict 100 computational steps ahead (Monte-Carlo steps, MCS), accelerating simulation evaluations by a factor of 562 times compared to single-core CPM code execution on CPU. Over short timescales of up to 3 recursive evaluations, or 300 MCS, our model captures the emergent behaviors demonstrated by the original Cellular-Potts model such as vessel sprouting, extension and anastomosis, and contraction of vascular lacunae. This approach demonstrates the potential for deep learning to serve as a step toward efficient surrogate models for CPM simulations, enabling faster evaluation of computationally expensive CPM simulations of biological processes.

This paper presents a TensorFlow-native implementation for automated crack detection in concrete structures, addressing the critical need for efficient and objective infrastructure monitoring. Leveraging a Convolutional Neural Network architecture with 24.8 million parameters, the model was trained on a large-scale dataset of 40,000 images, each with a 227x227 RGB resolution. The methodology, incorporating specific framework optimizations and a rigorous training configuration, achieved a remarkable overall classification accuracy of 99.375% on the validation dataset. The model demonstrated balanced performance with precision values of 0.993 and 0.994, recall values of 0.994 and 0.993, and F1-scores of 0.994 and 0.994 for both "No Crack" and "Crack" classes. This high accuracy, coupled with balanced metrics, underscores the model's effectiveness and reliability for practical applications. The proposed solution significantly enhances real-time structural health monitoring systems, mitigating the limitations of traditional manual inspections and facilitating proactive maintenance strategies for concrete infrastructure.

Abstract Deep learning has become a research focus in academia and industry due to its ability to effectively extract fault features from rotating machinery. However, given the variability of high-power variable-frequency industrial systems, existing models face the challenge of low accuracy in identifying electrical-erosion faults in rail transit motor bearings. Moreover, current models struggle to fully integrate the temporal information of such faults and exist technical challenges related to poor interpretability. To address the aforementioned shortcomings, this paper creatively develops a novel neural network architecture for global fusion of temporal sequence information, named BISR-Former, which focuses on solving the problem of difficult-to-identify bearing electro-corrosion faults in electric motors for rail transit in actual engineering applications. Firstly, inspired by the successful application of Transformer architectures in natural language processing, we make the first attempt to adapt i-Transformer to the task of motor-bearing fault diagnosis. We innovatively devise a Global Temporal Information Fusion module that comprehensively captures the global dependencies between long nonlinear sequences of motor-bearing data. By introducing this module into the proposed framework, the model gains the advantages of dynamic weighting and parallel computation. Secondly, recognizing the strong time-varying nature of the time-series data of bearing failures in rail transit electric motors, we have innovatively designed a bidirectional local time-series feature extraction module. By integrating this module into the proposed framework, it gains the ability to fuse bidirectional temporal modeling of motor bearings, enabling the framework to capture both long-range global dependencies and short-range local temporal features. Consequently, the framework attains a more comprehensive understanding of the sequential dynamics underlying motor-bearing fault evolution. Finally, extensive experiments on a real-world motor-bearing dataset confirm the proposed framework's superior performance and strong generalization capability. At the same, t-SNE was introduced into the proposed framework to enhance the interpretability of the fault-feature extraction process.

Alzheimer’s disease gradually erodes brain function, stringently disrupting memory andreasoning, expressly among older adults. Identifying the condition in its preliminary stages is decisive fortimely support and potentially more operative care.This study investigated the application of deep learningmodels for the automated detection of AD from MRI images. Three Convolutional Neural Network (CNN)architectures are utilized specifically—VGG16, Xception, and ResNet50. The models are evaluated in bothbinary classification and multi-class classification. Standard evaluation metrics are used to assess modelperformance. For binary classification, ResNet50 had the highest accuracy (97.96%), followed by VGG16 (97.10%) and Xception (95.93%). In multi-class classification, ResNet50 additionally led (95.39%), slightly ahead of VGG16 (94.92%) and Xception (94.93%).These results underscore the strong potential of ResNet50, in particular, for clinical application, demonstrating reliable generalization to previously unseen MRI images. The study highlights the potential of deep learning models to enhance early detection of Alzheimer’s disease by supporting clinical diagnosis, improving accuracy, and enabling timely interventions. Automated MRI analysis may also reduce costs and expand access to quality screening, especially in resource-limited settings reinforcing the growing case for integrating AI into medical imaging workflows.

Using lightweight convolutional neural network to identify ventilation/perfusion scintigraphy for acute pulmonary embolism.

Curriculum negative mining for temporal networks.

ABSTRACT Recent advancements in deep learning and the utilization of pre‐trained convolutional neural network (CNN) architectures have led to enhancements in classification tasks. However, these architectures often entail millions of training parameters, posing challenges for real‐world deployment. In this work, we propose an iterative Gaussian feature extractor with a custom 3‐layer CNN network (IGF‐CNN) coupled with a feedforward artificial neural network (ANN) classifier. The input images undergo pre‐processing before being fed to the proposed IGF‐CNN and then ANN classifies the input into Covid‐19, non‐Covid‐19 and pneumonia classes. The suggested model demands considerably fewer parameters and reduces training time substantially and achieves accuracies of 99.80%, 98.78%, 99.0%, respectively, across three different benchmark datasets. We have also performed cross‐dataset validation and obtained consistently good results, further demonstrating the robustness of the proposed approach. The proposed architecture is accurate and efficient and can be integrated with real‐time systems.

In-MEMS analog neural network architecture for signal denoising

NiNet: A new invertible neural network architecture more suitable for deep image hiding

A systematic review on automatic segmentation of renal tumors and cysts using various convolutional neural network architectures in radiological images.

Weakly supervised learning for scar reconstruction in personalized cardiac models: Integrating 2D MRI to 3D anatomical models.

Abstract Quantum-classical hybrid neural networks have recently emerged as a prominent research direction, combining the expressive power of quantum computing with the robustness and scalability of classical deep learning. These models have demonstrated significant potential, especially in image classification scenarios. However, existing hybrid architectures often fail to jointly model spatial and channel-wise features, which may constrain their representational capacity. In this work, we propose a novel quantum-enhanced convolutional neural network architecture that integrates Quantum Positional Encoding (QPE) and Quantum Channel Attention (QCA) mechanisms. These components respectively enhance spatial and channel-wise feature representations, while ensuring compatibility with conventional training pipelines. Experiments on the MNIST and Fashion-MNIST datasets demonstrate that the proposed model consistently outperforms classical CNN baselines in classification accuracy. Furthermore, experiments on the challenging CIFAR-10 dataset further validate the competitive performance of the proposed model, highlighting its strong generalization capability. These findings underscore the practical value of integrating quantum mechanisms into deep learning and emphasize their potential to advance intelligent visual recognition systems.

The versatile, Effective Stress Beta Method can be applied to driven pile design across various soil profiles. However, it is not widely practiced, potentially due to unreliable FHWA guidance for selecting pile design coefficients (β & Nt). This study addresses this issue by evaluating FHWA design coefficients against measured data from the Deep Foundation Load Test Database (DFLTD) V.2. The investigation reveals significant disparities between measured load-carrying capacities (Qm) and capacities calculated using FHWA design coefficients (Qc). The design coefficients, β & Nt, were then systematically back-calculated from load test data and were found to differ from the FHWA design coefficients, as hypothesized. In order to improve the design coefficients selection guidance, a neural network (NN) and machine learning (ML) based approach is proposed. The models BetaSPTNet (Artificial-NN for β) and NtSPTNet (Dense-NN for Nt) outperformed other models in predicting β & Nt. The study showcases NN’s adaptability in handling ambiguous correlations such as the one between geotechnical engineering properties of soil and β & Nt design coefficients. The proposed NN architecture improves precision and reduces uncertainty in determining β and Nt using geotechnical properties derived from SPT and CPT soil exploration data. Finally, we propose a modified approach integrating the traditional Beta design method with NN-predicted design coefficients. This integration significantly enhances the accuracy of calculated pile load-carrying capacities.