CNNs vs. transformers: A benchmark for multi-class marine debris identification.

Automated microplastic quantification is currently compromised by morphological mimicry: air bubbles, organic biofilms, and sediment create high false-positive rates in standard Convolutional Neural Networks (CNNs).This study introduces AquaEye, a containerized computer vision framework that mitigates artifact misclassification by fusing Bayesian Deep Learning with ISO-standard morphometrics.Unlike deterministic U-Net implementations, we deploy a Monte Carlo Dropout inference pipeline to estimate epistemic uncertainty, enabling the suppression of predictions where model variance exceeds a safety threshold.To enforce physical validity, a post-processing geometric gate rejects candidates based on Circularity (4A/P 2 ) and Solidity, filtering non-polymer structures that bypass the neural filter.The system, deployed via a Dockerized microservices architecture, ensures reproducibil-ity often absent in -lab-bench scripts.Experimental validation confirms that AquaEye statistically decouples true microplastic instances from background noise, offering a robust alternative to manual microscopy for highthroughput environmental monitoring.

Likelihood-free parameter inference for spatiotemporal stochastic biological models using neural posterior estimation.

Deep chemometrics with convolutional neural networks for the detection of honey adulteration using Fourier transform infrared spectroscopy.

ABSTRACT Passive Acoustic Monitoring (PAM) offers a powerful approach for detecting and assessing the presence of invasive species, thereby supporting the conservation of native ecosystems. In this study, we developed a species-specific classification model using convolutional neural networks (CNNs) to analyse the nocturnal calling activity patterns of Pelophylax nigromaculatus and Dryophytes leopardus in rice paddies, a microhabitat where interactions between translocated and native species are of ecological concern. Despite environmental noise from bird calls and wind, the mel-spectrogram-based model classified anuran vocalisations with high accuracy (88.45%). Misclassifications at dawn were mitigated by limiting the ecological analysis to specific nocturnal periods. The results revealed clearly distinct peak times of nocturnal calling activity between the two species at the study site. These findings demonstrate the usefulness of deep learning for describing fine-scale activity patterns in field-recorded amphibian soundscapes. This study provides methodological insights into the acoustic monitoring of domestically translocated and native species and highlights the potential of bioacoustics approaches for ecological assessment in paddy environments. Future research should focus on refining species classification models and integrating sound-source separation for more accurate species-specific assessments of calling activity patterns.

Explainable 3D VGG-style convolutional neural network for pediatric hydrocephalus detection on computed tomography: A segmentation-free and fully volumetric deep learning framework.

ABSTRACT Joint clustering of hyperspectral and Light Detection and Ranging (LiDAR) data is challenging due to their heterogeneity and differing spatial-spectral characteristics. To address this, we propose an adaptive multi-view graph convolutional network (MVGCN) that integrates visual Bidirectional Encoder Representations from Transformers (VisualBERT), referred to as MVGCN-VisualBERT, to extract high-level semantic features from both modalities. These features form a superpixel-level graph that preserves spatial structure while reducing redundancy. A multi-view graph convolutional network then propagates and aggregates information to enhance cluster cohesion. Evaluated on the MUUFL and UH2013 datasets, MVGCN-VisualBERT outperforms state-of-the-art methods, achieving improvements of 2.8% in overall accuracy, 2.6% in the Kappa coefficient, 2.9% in normalized mutual information and 5.8% in the adjusted Rand index on MUUFL. These results highlight the potential of the proposed approach for improving unsupervised multimodal land-cover analysis in remote sensing applications.

Integrating Contrast Colour Correction (CACC) and Convolutional Neural Networks (CNN) can help fish breeders in earlier classification and identification of Cryptocaryon fish disease (protozoan white spot disease). Disease identification accuracy is enhanced through such method by adaptive colour changes and CNN feature extracting ability, thereby boosting underwater image clarity. Unlike traditional rule-based systems that relies on expert knowledge despite being error-prone, existing methods focus on visual quality without classifying impact influence. Early disease identification is hampered in terms of efficiency due to machine learning methods reliant on abundant human expertise other than efficient feature extracting. An artificial intelligence (AI)-oriented computer model is introduced for existing research in overcoming limitations and eliminating subjectivity. The model employs a proprietary method of diagnosing fish disease through underwater images examination that yields objective outcome. Several CNN structures such as GoogleNet, ResNet-101, AlexNet, ResNet-50 as well as VGG-16 are tested on its performance. The current study shows integration of CACC with CNN through a set of 15000 images boosting up model performance in Cryptocaryon fish disease detection. The introduced novel method significantly enhances performance with 99.53% accuracy, 99.08% precision along with 100.00% recall. This efficient, accurate approach can significantly reduce the workload of experts and fish farmers while promoting sustainable aquaculture and healthier aquatic ecosystems. • Contrast-Adaptive Colour Correction (CACC) and Convolutional Neural Network (CNN) enhance fish images for accurate Cryptocaryon fish disease detection. • 6500 image datasets from National Fish Health Research aids AI training and testing. • One of the CNN architectures, ResNet50 outperforms others with 99.52 % accuracy. • 99 % accuracy across key metrics ensures early disease detection. • AI-driven approach supports sustainable aquaculture and reduces losses.

Despite the significant progress made in developing different in silico methodology for structure activity research over the past few decades. The ability to predict correlation structure activity (CSA) from absorption distribution metabolism excretion (ADME) descriptors to select indolizine compounds for human papilloma virus (HPV) anticancer activity continues to pose a challenge. This study employed five machine learning (ML) algorithms for classification, viz., stochastic gradient descent (SGD), random forest (RF), support vector machine (SVM), convolutional neural network (CNN), and logistic regression (LR), to perform the classification based on ADME-related physiochemical descriptors of 8,900 indolizine compounds to predict the CSA. The present study focuses on 26 well-known parameters to optimize the results, which are utilized for ML models SGD, RF, SVM, CNN, and LR for classification. The CNN achieved the best results with the highest overall accuracy and average loss values of 98.33% and 0.16, respectively. On the other hand, the SGD, RF, SVM, and LR recorded the accuracy values of 95.32%, 93.23%, 96.03%, 94.03%, and loss values of 0.046, 0.067, 0.039, and 0.059, respectively. It is stated that from the obtained results, the CNN is performing better compared to other methods. The cross-validation and results are done with the relationship of descriptors, viz., accuracy, correlation, distribution, area under the receiver operating characteristic, area under the precision recall curve, and bootstrap error analysis. This study demonstrated the utility of ML to facilitate early prediction of indolizine compounds for HPV anticancer activity in preclinical development.

Using neural networks to understand static and dynamic cues in facial expression recognition.

Practical guidance about tradeoff choices between accuracy, efficiency and deployment ability in deep convolutional neural network architectures for land use and land cover classification has been largely unavailable because each study evaluates architectures differently. This paper provides a controlled comparative assessment of four popular convolutional neural network architectures for land use and land cover classification visual geometry group19, ResNet50, Inception_Version3 and MobileNet_V2 using the Euro_SAT benchmark which includes 27,000 Sentinel2 red green blue images that have been cut into 10 land uses classes. The convolutional neural network architectures were all trained and evaluated through the same preprocessing, augmentation, data splitting, training procedure and metric as follows: Overall accuracy/F1 macro averaging class by class confusion matrix convergence dynamics efficiency metrics (number of parameters and inference-oriented considerations). The results indicate that modern architectures provide significantly better than older sequential baselines: ResNet50 provided the highest total accuracy (>97%), along with consistent convergence behavior; InceptionV3 improved discrimination for classes with both ambiguous visual appearances and linear structures (river, highway); MobileNetV2 was able to achieve high accuracy (>94%), but had an order of magnitude less number of parameters than the other architectures and is well suited to lower source or real time application scenarios. Finally, this paper maps convolutional neural network outputs into an energy transition decision workflow (Renewable Siting → Corridor Constraints → Monitoring), and demonstrates how land use and land cover layers derived from convolutional neural networks can support net zero resource planning. • Benchmarks 4 CNNs on EuroSAT using identical training and evaluation setup. • ResNet50 achieved highest accuracy (>97%) for LULC classification. • MobileNetV2 reduced parameters by ∼6× while maintaining >94% accuracy. • InceptionV3 improved corridor feature detection (roads, rivers). • Enables decision-grade LULC for CCUS, geothermal, and net-zero planning.

The rapid growth in global population necessitates efficient energy management solutions for sustainable living. Smart Building Energy Management Systems (SBEMS) play a crucial role in achieving this goal by leveraging automation and advanced analytics. This study proposes a novel Deep Learning and IoT-based SBEMS approach to predict energy consumption, classify buildings into energy-demand clusters, and optimize the monitoring and operation of electrical equipment. While traditional statistical methods have been widely used for load forecasting, recent advancements in deep learning provide robust alternatives to address the inherent complexity of nonlinear energy consumption patterns. This research employs regression analysis and state-of-the-art neural network architectures, including Single-Step and Multi-Step Dense Models, Convolutional Neural Networks (CNNs), and Long Short-Term Memory (LSTM) networks, to enhance prediction accuracy. Additionally, K-means clustering is introduced to segment buildings into distinct energy-demand categories, ensuring optimal energy utilization. Unlike prior studies that often lack a comprehensive approach, this work integrates all critical features under a unified framework. By applying these advanced methodologies to a unique dataset, the proposed system demonstrates improved accuracy in energy load forecasting and clustering, providing a significant contribution to the field of smart building energy management. The experimental results demonstrate that CNN and LSTM models significantly outperform conventional statistical approaches in capturing nonlinear energy consumption patterns. These outcomes support proactive energy scheduling, peak-demand mitigation, and scalable smart building energy management, offering practical value for facility managers, utility operators, and policymakers.

CDP-KDNet: Curriculum-Guided Dynamic Pruning and Knowledge Distillation for Resource-Efficient Ultrasound Elastography.

Alzheimer's disease (AD) is one of the most prevalent neurodegenerative disorders worldwide, requiring early identification for timely intervention and to slow disease progression. However, existing diagnostic approaches, while effective at later stages, remain limited in detecting early-stage AD. Handwriting analysis has recently emerged as a non-invasive, cost-effective, and ecologically valid digital behavioral biomarker that reflects neurocognitive impairment. This review examines the role of handwriting as a neurocognitive marker for AD, focusing on integrating deep learning methodologies to enhance early diagnostic accuracy. It also elucidates the neurocognitive mechanisms linking handwriting behavior and AD, addressing current methodological and translational challenges. We performed a PRISMA-informed structured literature search and narrative synthesis of handwriting- and drawing-based studies for detecting AD/mild cognitive impairment (MCI), including offline handwriting images and online pen-stroke kinematics captured by digital devices. Task paradigms, data dimensions, preprocessing pipelines, modeling strategies (traditional machine learning and deep learning), evaluation practices, and translational considerations were summarized, and studies were organized by detection purpose and analytic approach. Our findings show that handwriting-based models generally discriminate AD/MCI from healthy controls with accuracy exceeding 80%, while deep learning models (e.g., convolutional neural network and multimodal Transformer fusion) approach 90% in structured tasks like clock drawing and figure copying. Online kinematic markers (e.g., reduced velocity, prolonged in-air time, increased pausing, and pressure instability) recur across studies, and multimodal integration with speech, gait, or facial signals can further improve sensitivity and ecological validity, although most studies are small and single-center.

Enhancing breast mass detection: Super-resolution multi-spectral transmission imaging with unstructured sinusoidal illumination.

This study was to optimize the current methods for identifying and predicting the risk of critical illness in patients with connective tissue disease-associated interstitial lung disease (CTD-ILD). First, 200 patients diagnosed with CTD-ILD were included, and detailed demographic, serological, and imaging data were collected. Second, a risk identification and prediction framework was constructed based on multivariate logistic regression and machine learning algorithms (random forest (RF) and convolutional neural network (CNN)) to identify significant determinants of critical illness. Finally, the overall performance of each model was evaluated using K-fold cross-validation and external validation procedures. A feature ablation experiment was conducted based on the optimal random forest model to validate the independent contribution of each core predictor. The results showed that the logistic regression, random forest (RF), and CNN models were all successfully constructed and validated, among which the RF model demonstrated the best overall performance, with an accuracy of 85.7%, an area under the curve (AUC) of 0.88, a sensitivity of 83.5%, and a specificity of 88.2%. The ablation experiment confirmed that each feature had independent predictive value, with the most significant decline in model performance observed after the removal of IL‑6. Among them, the individual AUC value of interleukin-6 (IL-6) reached 0.981. Significant risk factors included patient age, C-reactive protein (CRP) level, presence of honeycomb lung on imaging, and the ratio of arterial oxygen partial pressure to inhaled oxygen concentration (PaO2/FiO2). The model in this study demonstrated satisfactory predictive ability and stability in both internal and external validation phases. The random forest model performed excellently in predicting the likelihood of critical illness in patients with CTD-ILD.

Image-based vulnerability detection based on a hybrid deep learning model in the Industrial Internet of Things using convolution neural network and transformer architectures

Seafood provides high-quality protein and essential nutrients but is highly susceptible to rapid postharvest deterioration. Conventional quality evaluation methods are often destructive, labor-intensive, and difficult to implement in real-time industrial and consumer settings. In recent years, deep learning-assisted computer vision (DL-CV) has emerged as a promising technical route for nondestructive and rapid seafood quality assessment in industrial processing lines and retail inspection systems. This review synthesizes recent advances in DL-CV for seafood quality evaluation from both technical and application-oriented perspectives. The technical framework is first clarified by comparing visible-light imaging with other imaging modalities and by discussing the transition from traditional machine learning to end-to-end deep learning-based feature extraction. Recent developments in representative deep learning architectures, such as convolutional neural networks and vision transformers, are then summarized alongside emerging trends toward lightweight design, architectural enhancement, and interpretability. Current application scenarios are further reviewed systematically, with particular emphasis on freshness evaluation, including key image regions and two dominant labeling strategies: storage time-based labeling (STBL) and traditional indicator-based labeling (TIBL). Additional applications such as species identification, weight estimation, defect detection, and quantitative determination of compositional attributes are also discussed. Overall, DL-CV demonstrates strong potential for accurate, nondestructive seafood quality prediction, especially when leveraging visible-light imaging systems that are easily deployable in practical environments. Future research should focus on finer quality differentiation, multidimensional and integrated quality evaluation, and scalable deployment in both industrial processing and consumer-oriented applications.

To find the cause of infectious diseases such as urinary tract infections, a the diagnostic patterning test was used. In this method, the whole blood and serum samples for every person were dried on a glass substrate to create a special pattern due to formation of central and peripheral regions. This process was performed during 1h at 25°C which is much shorter than the classical and instrumental methods. The resulting patterns can be observed by an optical microscope. The study aimed to classify the UTI patients and healthy controls (N=600). In addition, the images of dried patterns were captured by a camera and the classification analysis was done by convolutional neural network algorithm, executed by graphical user interface. The result of analysis was available after a few minutes, indicating that the proposed method achieved accuracies of 90.8% (through the analysis of the dried blood spot) and 86.6% (through the evaluation of the dried serum spot) for discriminating the patients from healthy participants. This method could become popular because it has a simple design and does not require skilled personnel, toxic reagents and expensive equipment. Most importantly, it can identify the contaminated sample as well as the contaminant in a shorter time than other common methods.

Vision Transformer (ViT) has shown impressive performance in image restoration due to its ability to capture a large receptive field. However, its complexity grows quadratically with input resolution, limiting its applicability for high-resolution images. In contrast, Convolutional Neural Networks (CNNs) are computationally efficient but are constrained by their inherently local receptive fields, which limit their ability to capture long-range pixel relationships. To address these challenges, we propose StarIR, which possesses the efficiency of CNNs while also capturing a large receptive field, similar to Transformers. StarIR incorporates two key innovations: 1) a dual-domain representation learning framework, with one branch processing spatial details and the other focusing on mesoscale interactions in the frequency domain; and 2) a high-dimensional feature fusion mechanism, the Star operation, which fuses information from both domains through element-wise multiplication, thereby enhancing representational capacity without increasing network width and depth. Our Star operation is followed by a channel attention unit to facilitate global feature modeling and enhance channel-wise interactions. Building on our straightforward yet powerful design principles, StarIR achieves state-of-the-art performance across 21 datasets covering six single-degradation image restoration tasks. Furthermore, our model performs favorably against leading algorithms in two all-in-one settings and demonstrates robustness on two composite-degradation datasets. In addition, StarIR extends well to several domain-specific applications, including ultra-high-definition (UHD) imaging, remote sensing, medical imaging, and underwater image enhancement.