Search Results (1,440)

Background/Objectives: Tissue fold artifacts observed in hematoxylin-eosin (H&E)- stained preparations used in histology education can complicate the assessment of normal tissue architecture and affect students’ accurate interpretation of microscopic structures. This study aimed to automatically detect and localize tissue fold artifacts in histology teaching preparations using deep learning methods. Methods: A total of 2127 hematoxylin and eosin (H&E)-stained histological images of brain, kidney, liver, small intestine, and testis tissues obtained at 10× magnification were used in the study. The dataset consisted of 899 clean/artifact-free images and 1228 images containing tissue fold artifacts. Seven deep learning architectures, including convolutional neural networks (CNN)-based models and a vision transformer-based model, were evaluated for image-level classification: ResNet18, ResNet50, DenseNet121, EfficientNet-B0, EfficientNet-B3, ConvNeXt-Tiny, and Swin-Tiny. Classification performance was evaluated using an organ-based testing approach, a slide-level train–test split in which images from the same histological slide group were retained within a single subset, a general image-level 80/20 train–test split, and five-fold cross-validation. A DeepLabV3-ResNet50-based segmentation model was trained using QuPath-prepared masks to determine fold regions at the pixel level. Grad-CAM was used for qualitative visualization and quantitative comparison with manually annotated fold regions. Results: All classification models showed excellent performance overall. The Swin-Tiny model was the most accurate in terms of general image-level classification, scoring 99.06% for accuracy, 99.19% for F1-score, and 99.98% for AUC. In slide-level classification, EfficientNet-B3 showed the best performance in terms of accuracy (99.54%) and AUC (99.99%). No pairwise statistical difference was found among the models using Holm adjustment. In the five-fold cross-validation setting, ResNet50 achieved the best performance with the accuracy of 98.73 ± 0.54%. In the detailed small-intestine error analysis, classification accuracy was 67.67%, with high sensitivity (98.59%) but low specificity (24.34%), mainly because of false-positive predictions. In the segmentation evaluation, the final DeepLabV3-ResNet50 model trained with combined BCE + Dice loss resulted in Dice of 0.7630 ± 0.2425 and IoU of 0.6661 ± 0.2577 on the independent test set. False positive segmentations were rare in 899 artifact-free images (only 0.33% of images had a tissue fold region of at least 1%). The quantitative Grad-CAM analysis showed poor spatial agreement with the manually annotated tissue fold masks (Dice = 0.2423; IoU = 0.1441). In the independent MPP10 dataset, ResNet50 achieved 89.63% accuracy and 88.49% F1-score. Conclusions: The suggested method was able to detect tissue fold artifacts as well as localize them in the teaching images of histology. The strong performance recorded at the slide level validates the internal findings, but the limited performance in the case of small intestine images and on the external dataset reveals that tissue structure remains a crucial factor. Full article

Scoliosis is a complex three-dimensional spinal deformity that requires early detection, accurate radiological assessment, individualized treatment planning, and long-term monitoring. In recent years, artificial intelligence (AI) has emerged as a promising tool across multiple stages of scoliosis management, including screening, automated imaging analysis, deformity classification, prediction of progression, surgical planning, postoperative outcome assessment, and patient education. This narrative review summarizes current and emerging applications of AI in scoliosis, with particular emphasis on studies published after 2023. Deep learning algorithms, including convolutional neural networks, U-Net-based architectures, transformer models, and generative approaches, have demonstrated high accuracy in automated Cobb angle measurement, vertebral segmentation, coronal and sagittal parameter assessment, and radiation-free screening using surface topography or smartphone-based photographs. Machine learning models have also shown potential in predicting curve progression, treatment response, risk of postoperative complications, and patient-reported outcomes by integrating radiological, clinical, biomechanical, and, increasingly, multimodal data. In parallel, large language models and generative AI tools are being investigated for patient education, communication support, readability improvement, and research hypothesis generation. Despite these advances, important limitations remain, including limited external validation, dataset heterogeneity, potential algorithmic bias, insufficient interpretability, and incomplete integration into clinical workflows. Moreover, while AI systems show strong performance in automated measurement and screening tasks, their role in complex therapeutic decision-making, such as Lenke classification, fusion-level selection, and autonomous surgical planning, remains experimental. Overall, AI has the potential to improve the precision, efficiency, and personalization of scoliosis care; however, prospective multicentre studies, transparent reporting, explainable model design, and regulatory validation are essential before widespread clinical implementation. Full article

Deep learning has shown promise for cotton leaf disease detection, yet two critical gaps remain. First, most studies rely on a single model (Convolutional Neural Network-CNN or Transformer) and do not explore how to effectively fuse these complementary architectures. Second, eXplainable AI (XAI) methods are often used qualitatively, lacking objective benchmarks to guide method selection. To address these gaps, we evaluate six backbone models, comprising four CNNs (ResNet50, EfficientNet-B0, DenseNet121, and MobileNetV2) and two Vision Transformers (ViT-Base and DeiT-Small), on the Kaggle cotton leaf disease dataset, which contains 1711 images across four classes. We then systematically investigate five CNN–Transformer fusion strategies, namely concatenation, attention, weighted, ensemble, and variance-based fusion, to identify the most effective approach for disease classification. The best-performing individual models are DenseNet121 (92.40% accuracy) and ViT-Base (96.49% accuracy). Classification metrics include accuracy, balanced accuracy, precision/recall, F1-score, Cohen’s kappa, MCC, AUC, bootstrap confidence intervals, and McNemar tests. Computational efficiency (FLOPs, inference time, model size) is also reported. Concatenation fusion achieves the highest performance (accuracy = 99.42%, 95% CI: 98.2–100%, weighted F1 = 0.994, MCC = 0.992). For explainability, we quantitatively compare six XAI techniques, GradCAM, GradCAM++, ScoreCAM, LayerCAM, EigenCAM, and AblationCAM, using the pointing game, IoU, AUC, and localization accuracy. EigenCAM yields the best overall explainability score. This study demonstrates that simple feature concatenation between dual backbones (CNN + Transformer) is highly effective for cotton leaf disease detection and provides a benchmark for XAI method selection in plant pathology. Full article

Nitrogen is a major determinant of crop growth, yield formation, and grain quality, making precise nitrogen (N) management essential for sustainable production. This study integrated field canopy spectra acquired with an ASD FieldSpec 4 spectrometer and unmanned aerial vehicle (UAV)-based RGB imagery to estimate leaf nitrogen content in broomcorn millet across four growth stages. Vegetation indices (VIs) and gray-level co-occurrence matrix texture features were derived, and six algorithms—partial least squares (PLS), support vector machine (SVM), random forest (RF), one-dimensional convolutional neural network (CNN1D), one-dimensional residual network (ResNet1D), and one-dimensional U-Net (U-Net)—were evaluated. Among the spectral preprocessing methods, the first-derivative transformation showed the strongest relationship with leaf nitrogen content (r = −0.85). Models based on multi-source feature fusion consistently outperformed those using a single data source. The U-Net model using ASD + RGB + VIs + texture features achieved the highest test-set accuracy (R² = 0.913, RMSE = 1.851, and RPD = 3.382). SHapley Additive exPlanations (SHAP) analysis identified blue-band correlation (B_Correlation) as the most influential individual predictor (mean absolute SHAP value = 0.39), while texture features accounted for 45.170% of the cumulative feature importance. These results demonstrate the potential of multi-source remote sensing and deep learning for rapid, non-destructive assessment of leaf nitrogen content and precision nitrogen management in broomcorn millet. Full article

Background: Image splicing forgery detection is one of the most challenging problems in the field of image forensics as it involves identifying and localizing suspicious regions that are created by integrating contents from one or more different sources. The accurate detection and classification of splicing forgery still remains a difficult task because of the existence of overlapping image regions, which makes it complex to differentiate authentic and tampered images. This overlap causes a lack of feature representation, making it difficult to precisely detect the tampering in images. Also, the decision-making process of the model is often a black box, which makes it challenging to interpret and understand the rationale behind its decisions. Methods: To address these challenges, a Convolutional Block Attention Module (CBAM)–U-Net with Cross-Connected Filters–Convolutional Neural Network (CCF-CNN) is proposed to achieve precise detection and localization of spliced regions. The CBAM enhances spatial and channel-wise attention, enabling accurate localization of forged regions. The dual-phase CCF-CNN is incorporated with cross-connected filters to differentiate between the authentic and tampered regions by extracting global and local features. Additionally, a surrogate heatmap mechanism is introduced using intermediate decoder features to generate patch-level visual explanations, enabling precise localization of the spliced regions, thereby improving the model’s transparency in decision-making. Results: The proposed CCF-CNN obtains a high accuracy of 99.84% on the CASIA 2.0 dataset and an accuracy of 95.63% on the MISD. Conclusions: Compared to traditional CNNs such as VGG, ResNet and attention-based interpretability algorithms, the proposed model obtains higher performance in terms of detection and interpretability. Full article

Gradient-based inversion methods are widely used for time-domain airborne electromagnetic (AEM) data, but their results are commonly affected by the initial model and regularization-induced smoothing. These limitations are particularly evident when thin layers or alternating high- and low-resistivity structures need to be resolved. To address this problem, we propose a gradient-guided iterative enhancement (GGIE) inversion that combines a limited-iteration Gauss–Newton (GN) inversion with a lightweight U-Net. In each GGIE inversion iteration, the GN module first produces a coarse inverted model (IM) that preserves the main data-driven geoelectric trend but is still affected by regularization-induced smoothing. The trained U-Net then predicts a structurally enhanced model (PM) from the observed data and the IM. A data-misfit-guided adaptive approach is proposed to calculate the weight coefficients of the IM and PM and to construct an update model (UM). These coefficients are further smoothed by a momentum term so that the relative contributions of the IM and the PM are adjusted adaptively during the iterations. This design reduces error propagation from either component alone and dynamically balances learned structural enhancement with physics-based data consistency. The UM then serves as the initial model for the subsequent GN inversion. GGIE inversion is tested on synthetic data, and the results show that it is most beneficial for complex multilayer structures, for which it reduces the mean relative error and root mean squared error (RMSE) by 49.0% and 28.6%, respectively. Compared with the physics-informed neural network (PINN) baseline, GGIE inversion reduces the model relative error, log-domain RMSE, and data misfit by 19.4%, 7.0%, and 71.7%, respectively. Moreover, compared with U-Net alone, GGIE inversion reduces the data misfit by 85.4%. The proposed method is further applied to field data acquired from the Fox River area in Wisconsin, USA. The main advantage of GGIE inversion is its ability to resolve complex multilayered structures, thin layers, and sharp resistivity contrasts with improved accuracy and stability. Full article

Accurate segmentation of skin lesions and gastrointestinal polyps is essential for early diagnosis and treatment planning. Currently, Convolutional Neural Networks (CNNs) are limited by local receptive fields, missing small lesions. While Transformers model global context, their quadratic computational complexity incurs high costs. To address these limitations, we propose the Wavelet–Vision Mamba UNet (WVM-UNet), integrating State Space Models (SSMs) for linear-complexity long-range dependencies and wavelet transforms for fine-grained feature extraction. The network employs a Wavelet-based Residual State Space (WRSS) block, combining the multi-scale decomposition of discrete wavelet transforms with Vision Mamba to efficiently capture global features. A Fused Channel–Spatial Attention (FCSA) mechanism is incorporated to adaptively recalibrate feature representations. Additionally, we construct an Encoder–Decoder Semantic Connection (EDSC) to replace traditional skip connections, effectively bridging the semantic gap between cross-level features. Experimental results on multiple public datasets demonstrate the competitive performance of our method. Specifically, on the ISIC 2017 dataset, WVM-UNet achieves an mIoU of 82.94% and a DSC of 90.67%, outperforming the Mamba-based VM-UNet by 2.71% in mIoU. These results indicate our architecture effectively captures discriminative features for precise medical image segmentation. Full article

The angular distribution of surface-strands in oriented strand board (OSB) is closely associated with board mechanical properties and mat formation quality. By acquiring surface images through vision sensing and combining them with deep learning-based segmentation, the angle classes of OSB surface-strands can be segmented and statistically analyzed automatically. However, OSB surface images contain complex strand textures, blurred boundaries, local adhesion between adjacent strands, and subtle differences among neighboring angle classes. To address these challenges, this study proposes a direction-aware dual-branch semantic segmentation network (DiBiNet) for pixel-level segmentation of surface-strand angle classes. OSB surface images were collected using a Hikrobot MV-CE120-10UC color industrial camera, and an 11-class dataset was constructed, including the background and ten angle classes from 0° to 90°. The samples were cropped to 512 × 512 pixels, and an improved angle-semantic-consistent Copy–Paste strategy was used to augment the training data. DiBiNet enhances directional feature representation through a Directional Strip Detail Enhancement Module, improves semantic feature modeling by combining MobileNetV3-Small with a DS-MobileViT Block, and fuses the two branches through a Bilateral Gated Fusion Module. Considering the continuity among angle classes, Direction Vector Auxiliary Supervision is introduced to map discrete angle labels into continuous direction vectors, thereby improving discrimination among neighboring classes. Experiments on the self-constructed dataset show that DiBiNet achieves a mean Intersection over Union (mIoU) of 0.8532, an overall pixel accuracy (Acc) of 0.8823, and a Dice coefficient of 0.8623, outperforming several representative semantic segmentation models. After 8-bit integer (INT8) + 16-bit floating-point (FP16) mixed quantization, the model achieves a neural processing unit (NPU) inference speed of 34.0 frames per second (FPS) on the RK3588 platform, demonstrating its potential for vision-based sensing and edge AI inspection. Full article

Ischaemic stroke constitutes a leading cause of global disability. Rapid extraction of the infarct core from multimodal computed tomography perfusion (CTP) imaging guides reperfusion therapy and clinical decision-making. Deep learning algorithms automate this delineation, yet hospital translation is hindered by high-dimensional data, inter-scanner variability, and the low contrast of early ischaemia. Architectural comparisons in the literature frequently carry methodological biases originating from disparate preprocessing protocols and data partitions. This study reduces these variables by evaluating three segmentation strategies under a shared preprocessing pipeline and an identical data partition using the ISLES 2024 dataset. Three models were trained on the same 133-patient partition using a shared preprocessing pipeline based on morphological skull-stripping and modality-specific clinical intensity ranges. The data, preprocessing, and partitions are held constant across models, while framework-dependent factors (optimiser, patch size, physical field of view, spatial resampling, augmentation policy, and model capacity) remain coupled to each architecture and are therefore treated as part of the compared strategy rather than as fully isolated variables. The first of these is a single-stage 5-channel nnU-Net, followed by a two-stage cascaded nnU-Net (2 and 7 channels) and a lightweight Transformer (SegFormer3D). Evaluation on a fixed 15-patient held-out test set isolated the architectural performance. The cascade model achieved the highest Dice Similarity Coefficient (0.224). The single-stage nnU-Net provided the most precise volumetric estimation, recording an Absolute Volume Difference (AVD) of 23.70 mL and a lesion-wise F1-score of 7.60%. On the other hand, SegFormer3D returned the lowest overall metrics (DSC 0.163, AVD 27.28 mL, F1 2.30%). In the small held-out cohort, paired statistical testing did not reveal significant differences between models, so the reported orderings describe the present dataset and experimental configuration rather than a general architectural law. Within these limits, the local inductive bias of the convolutional models retained an empirical advantage over the single Transformer evaluated when processing this moderately sized neuroimaging dataset, and complex cascade topologies offered only marginal gains compared with a well-calibrated single-stage network. Although the predictive segmentation of infarcted tissue at acute stages still demands computational improvements, these results suggest that preprocessing quality is at least as decisive for clinical impact as increasing the complexity of neural architectures. Full article

As a fundamental technique in geophysical exploration, gravity inversion plays a crucial role in geological structure interpretation and mineral resource assessment. Recent advances in deep learning, particularly the remarkable capabilities of convolutional neural networks (CNNs) in image recognition, object detection, and semantic segmentation, have provided innovative solutions to nonlinear inverse problems in geophysics. However, conventional data-driven approaches often yield physically unrealistic inversion results due to the inherent non-uniqueness of solutions. This study proposes a novel 3D deep-learning gravity inversion framework incorporating well-logging prior constraints, which establishes a density-position relationship-based constraint mechanism through synergistic integration of surface gravity anomaly data and downhole petrophysical parameters. We develop a new neural network architecture that incorporates well-logging data as hard constraints for gravity inversion while preserving both density characteristics and spatial information of subsurface formations. Significantly, we implement the Convolutional Block Attention Module (CBAM) during network optimization, enabling selective enhancement of lithological features during backpropagation. This attention-guided mechanism achieves robust coupling between potential field anomalies and petrophysical parameters from well logs, substantially improving the spatial accuracy of 3D density reconstruction. Numerical experiments demonstrate that our method can accurately recover the density distribution of subsurface ore bodies. Comparative results show that the integration of well-logging information significantly enhances both solution reliability and structural consistency when compared to purely surface data-driven approaches. In practical application to field data from the San Nicolas sulfide deposit in Mexico, the proposed method outperforms conventional UNet-based approaches in terms of inversion performance. Full article

Accurate binarization of phase-detection probe signals (gas vs. liquid) is necessary for the estimation of local void fraction, interfacial velocity, and bubble statistics in gas–liquid flows, particularly in nuclear thermal–hydraulic experiments. Classical threshold-based methods—single or double level—perform well on clean laboratory signals but degrade under realistic industrial conditions where noise, baseline drift, and clustered (slug-like) events challenge fixed rules. This work investigates whether deep learning (DL) models trained exclusively on synthetic data can deliver robust, generalizable binarization on real probe measurements. We (i) build a parametric generator of realistic time series from bubbly pulse templates, extended to clusters/slug patterns and perturbed with controlled noise, drift, and oscillatory baselines; (ii) train four lightweight DL architectures—one-dimensional U-Net (UNET-1D), Temporal Convolutional Network (TCN), a minimal one-dimensional Convolutional Neural Network (CNN-1D), and a Bidirectional Long-Short Memory network (BiLSTM)—only on synthetic signals; and (iii) evaluate them against classical threshold methods using event-level and sample-level metrics. On synthetic signal evaluation, UNET-1D and TCN achieve near-perfect event detection and sub-millisecond onset errors. On real bubbly and slug flow sensor data, classical threshold-based methods remain highly competitive on clean sensor signals, while DL models retain advantages under non-stationary baselines and clustered events, yielding accurate void and timing with no hand-tuned assumptions. Results support DL as a practical, data-driven complement to fixed algorithms, particularly in noisy or drift-dominated measuring conditions typical of nuclear thermal–hydraulic loops and safety-relevant test facilities. Full article

Quantitative precipitation estimation (QPE) is a fundamental task of hydrometeorological applications, ranging from flash-flood detection to water-resource management. While weather radars offer superior spatial coverage compared with rain gauges, traditional estimation based on the empirical $Z-R$ (reflectivity factor vs. rainfall rate) relationship fails to capture spatial variability and tends to underestimate extreme rainfall. The idea pursued here is to learn the radar-to-rainfall mapping by means of a convolutional neural network (CNN) trained on co-located radar and rain-gauge data; thus, once trained, the network can convert radar reflectivity measures into rainfall values, even where gauges are unavailable. Although machine learning techniques are promising for this task, they typically demand large training sets. To operate in a limited-data regime, we introduce a U-Net architecture that separates the analysis of space from that of time: each block first looks at the structure of the reflectivity field and then at how it changes over consecutive radar scans, extracting spatiotemporal features from volumetric data with fewer parameters than a full three-dimensional filter. The model is evaluated on a severe convective event that affected Tuscany, Italy, on 2 November 2023, benchmarking its performance against the classical Joss–Waldvogel $Z-R$ relationship (suitable for convective events), a data-driven log-regression of weather-radar and rain-gauge data, and a baseline CNN architecture. The main advantages are negative bias—i.e., underestimation of rainfall—more than halved and correlation with rain-gauge measures more than doubled, under the same operational conditions. What is noteworthy is the capability of learning the model from radar and rainfall data taken in different times and places, as well as the possibility of converting a reflectivity map into a rainfall map without the need for simultaneous rain-gauge measures. This is an asset of fixed parametric methods; however, they are far less accurate. Full article

Left atrium segmentation from magnetic resonance imaging (MRI) is essential for ablation planning in atrial fibrillation; however, clinical MRI quality is often degraded by noise, artifacts, and incomplete spatial coverage, making traditional recurrent neural networks (RNNs) vulnerable to such distortions. We developed a hybrid architecture inspired by cortico–cerebellar interactions to enhance segmentation stability without compromising mean accuracy. We utilized the open ATRIA dataset (100 patients, isotropic 3D MRI scans with manual left atrium annotations). The model comprises a convolutional encoder, a cortical RNN, and a cerebellar predictive module trained to predict future encoder features across multiple temporal horizons, generating a corrective feedback signal for the RNN. Experiments were conducted on unperturbed and degraded datasets with performance evaluated using the Dice coefficient. On unperturbed data, the cerebellar model achieved a mean best Dice of 0.835 ± 0.032 vs. 0.832 ± 0.027 for the baseline. Under degraded conditions, it showed significantly higher Dice (0.815 ± 0.019 vs. 0.801 ± 0.021; p = 0.014) and Surface Dice (p = 0.040), with a directionally lower between-run variance, though this difference in variance was not formally tested given the limited number of runs. nnU-Net achieved higher absolute accuracy but required three orders of magnitude more inference time and an order of magnitude more parameters. The cerebellar module improved boundary accuracy and reproducibility relative to the non-predictive baseline at a fraction of nnU-Net’s computational cost, offering a lightweight alternative for settings where deploying a full 3D self-configuring model is impractical. Full article

Road-network extraction from very high-resolution (VHR) remote-sensing imagery remains a challenging task owing to the structural sparsity, topological complexity, and severe occlusions of road networks. Conventional graph-based approaches preserve topological consistency yet incur considerable computational overhead, whereas prevailing convolutional neural network (CNN) and Transformer architectures struggle to reconcile long-range contextual modeling with computational efficiency. To address these limitations, this study proposes RFM-UNet, a hybrid frequency and state–space network designed for road-network segmentation. Specifically, the encoder integrates Mamba blocks with an Anisotropic Directional Attention (ADA) module to jointly capture local geometric cues and global dependencies at linear computational complexity. In addition, a Multi-Scale Adaptive Fusion Module (MAFM) is introduced to dynamically recalibrate multi-stage features, thereby suppressing cross-scale interference and preserving the connectivity of narrow roads. To enhance robustness against shadow-induced occlusions, a Dual-Spectrum Aggregation Module (DualSpec) decouples the phase and amplitude spectra in the frequency domain and fuses them with spatial features, effectively mitigating spurious responses and background noise characterized by similar textures. Quantitative and qualitative experiments on three public datasets demonstrate that RFM-UNet consistently outperforms current state-of-the-art methods. Full article

Remote sensing images are frequently degraded by occlusions and missing observations, which significantly affect subsequent interpretation and analysis. Matrix completion provides an effective solution for recovering incomplete data; however, existing deep learning-based approaches often rely on random initialization, resulting in slow optimization and limited reconstruction quality under severe missing conditions. To address these issues, this paper proposes a two-stage neural network-based matrix completion framework that combines SVD-guided low-rank modeling with convolutional feature learning. Specifically, truncated singular value decomposition (SVD) is first employed to initialize the network and provide a coarse reconstruction by jointly modeling the global low-rank structure and nonlinear image representations. A U-Net-based convolutional autoencoder is then used to refine the reconstruction by exploiting local spatial correlations and multi-scale features. In addition, a channel aggregation strategy is introduced to improve structural consistency for multi-channel remote sensing images. The proposed framework adopts a training-data-free optimization paradigm, eliminating the need for external training datasets by optimizing the network parameters directly for each input image. Experimental results on synthetic and real remote sensing images demonstrate that the proposed method consistently outperforms conventional matrix completion methods and achieves competitive performance compared with recent deep learning approaches, particularly under random missing patterns and high missing-rate scenarios. Full article