Search Results (63)

During horizontal drilling in coal-rock gas reservoirs, the particle size distribution (PSD) of drilled cuttings directly affects drilling efficiency, hole cleaning, and wellbore stability. However, the evolution of cuttings PSD and its controlling mechanisms during coal-rock fragmentation remain insufficiently understood. In this study, a drill bit–coal-rock interaction model was established using the discrete element method (DEM) and calibrated against uniaxial compression experiments. The effects of weight on bit (WOB), rotational speed, and depth of cut (DOC) on cuttings PSD were quantitatively investigated. The results show that the relative influence on the maximum cutting size followed the order of DOC > WOB > rotational speed, whereas the influence on the average cutting size followed the order of rotational speed > WOB > DOC. Increasing DOC from 0.5 mm to 1.5 mm increased the maximum cutting size from 11.6 mm to 29.4 mm. Increasing WOB promoted the generation of medium- and large-sized cuttings, thereby increasing hole-cleaning requirements. Meanwhile, increasing rotational speed from 40 rpm to 90 rpm reduced the average cutting size and shifted the dominant cutting fraction from 4–6 mm to 1–4 mm. DEM observations reveal that cutting PSD evolution is jointly controlled by primary brittle fracture and secondary particle breakage through a five-stage fragmentation process involving stress concentration, microcrack initiation, crack propagation and coalescence, fragment detachment, and secondary fragmentation. Field validation using 146 cutting samples demonstrated the applicability of the proposed optimization strategy. Under the investigated drilling conditions, a DOC of approximately 0.5 mm and a rotational speed of 70–90 rpm were found to effectively limit oversized cutting generation. These findings improve the mechanistic understanding of cutting PSD evolution and provide practical guidance for drilling parameter optimization and hole-cleaning management in coal-rock gas horizontal wells. Full article

In Optical Camera Communication (OCC), precise localization of LED arrays under complex tilt conditions is a core challenge for reliable decoding. This paper proposes an OCC reception scheme for RGB-LED arrays that integrates YOLO-OBB rotated object detection with two-stage geometric correction. The system first employs a YOLOv8n-OBB model to extract a quadrilateral region of interest that tightly encloses the LED array boundary. This effectively suppresses background interference caused by superimposed perspective tilt and in-plane rotation. A coarse-to-fine two-stage correction framework is then applied. The first stage rapidly eliminates the dominant perspective distortion based on the detected bounding-box corners. The second stage performs a refined correction using the actual LED center positions. Two homography matrices are cascaded into a combined transformation, achieving two-stage correction accuracy through a single coordinate mapping. In the corrected image, K-Means clustering constructs a 16 × 16 LED topological grid. A locking strategy is adopted so that subsequent frames skip repeated LED detection and clustering. The steady-state per-frame processing time is reduced to approximately 78.9 ms. Experiments covered 16 cross-combinations of vertical tilt from 0° to 45° (0°, 15°, 30°, 45°) and in-plane rotation from 0° to 40° (0°, 15°, 30°, 40°). The uncorrected scheme and the horizontal-box scheme experienced severe bit errors or complete failure under complicated distortion. The proposed scheme maintained error-free transmission under all 16 tested conditions. The ratios of opposite sides of the corrected LED grid remained stable between 0.997 and 1.004. The system simultaneously achieves high reliability and low-latency real-time processing under complex geometric distortions. Full article

Localized failures of structural components can lead to serious social, economic, and environmental consequences, such as the collapse of an entire structure or part of it. Therefore, it is important to thoroughly investigate and justify the acceptance criteria for these components, taking into account their performance in extreme conditions. However, the scientific literature lacks a systematic analysis of how various factors can affect the resistance of structures and influence acceptance criteria under extreme conditions. Therefore, this study investigates the typical substructures of reinforced concrete frame buildings in areas that are potentially prone to local collapse. To assess their resistance and structural robustness, an analytical model has been developed. The results of 22 tests on typical substructures of monolithic and precast frames, reported in various research studies, were used to validate this model. Further, this analytical model was used to conduct a parametric study on the impact of various factors on the performance of substructures under extreme conditions. These factors included the depth-to-span ratio of the beam, the strength of the bond between the steel reinforcement and the concrete, the stiffness of the horizontal bracing within the substructure, and the proportion of the effective depth to the total depth of the beam section. It has been found that the ultimate rotation angle in the plastic hinge of beams increases as the ratio of the beam’s cross-sectional depth to the span increases. An increase in the bond strength between the reinforcement and concrete leads to a decrease in the ultimate rotation angles in the plastic hinge at the flexural and arch stages of resistance and, in some cases, to reinforcement rupture without transitioning to the catenary stage of resistance. A decrease in the ratio of the effective depth of the beam section to its overall depth leads to an increase in the load-bearing capacity at the catenary stage of 19%. Full article

Intelligent perception of watermelon targets in complex field environments is a key prerequisite for automated harvesting and future collaborative decision-making with agricultural large models. To address severe leaf occlusion, large pose variation, dense adhesion among adjacent fruits, and the inability of conventional horizontal bounding boxes to accurately represent target orientation under natural cultivation conditions, this paper proposes an improved YOLO11n-OBB-based method for rotated watermelon detection. During data preparation, a semi-automatic annotation strategy combining segmentation-mask assistance with circumscribed rectangle fitting was adopted to efficiently construct a watermelon OBB dataset that closely matches the true physical boundaries of the fruits. On this basis, three structural improvements were introduced to the YOLO11n-OBB baseline: an LSK module was selectively embedded into the middle and later stages of the backbone to enhance adaptive receptive-field modeling and occlusion reasoning in complex bac kgrounds; the original neck structure was replaced with a lightweight BiFPN to strengthen bidirectional feature fusion for targets with large-scale variation in field scenes; and KFIoU Loss was incorporated into the rotated box regression branch to alleviate angle sensitivity and boundary discontinuity, thereby improving the convergence stability of orientation parameter learning. On the constructed watermelon OBB test set, the improved model raised mAP@0.5 (OBB) from 0.871 to 0.931, mAP@0.5:0.95 (OBB) from 0.670 to 0.736, Precision from 0.885 to 0.931, and Recall from 0.849 to 0.908 relative to the YOLO11n-OBB baseline (relative gains of 6.89%, 9.85%, 5.20%, and 6.95%, respectively), while keeping the inference speed at 100 FPS and the parameter count at only 2.71 M. While maintaining a compact model size and high real-time performance, the proposed method significantly improved rotated detection accuracy in crowded and overlapping scenes. In addition, the detection results were encapsulated into a structured JSON perception interface, preliminarily demonstrating the integration pathway of this lightweight front-end for task planning and human–machine collaborative operations with agricultural large models, and indicating its potential for future intelligent agricultural decision-making. Full article

The rapid development of synthetic aperture radar (SAR) imaging technology has significantly enhanced maritime monitoring capabilities; however, SAR ship detection remains constrained by the limited scale and representation capacity of existing rotated bounding box datasets. Most publicly available datasets rely on horizontal annotations, which introduce redundancy and localization ambiguity in densely distributed and nearshore scenarios. Although rotated bounding boxes provide more precise geometric representation, large-scale multi-source rotated SAR datasets are still insufficient to support robust model training. To address this limitation, we construct a large-scale multi-source rotated SAR ship dataset (LMRD) consisting of 13,024 high-resolution image chips with over 38,000 annotated ship instances, covering multiple satellite sources, polarization modes, and diverse maritime environments, including offshore, nearshore, complex coastal, and densely distributed port scenes, thereby enhancing scene diversity and annotation precision. Furthermore, independent of the dataset construction, we propose a multi-domain feature fusion (MDF) framework built upon Oriented RCNN, which integrates high-frequency information and visual saliency cues to improve feature representation under complex backgrounds. Experimental results on the LMRD demonstrate that, compared with the baseline Oriented RCNN, the proposed MDF framework achieves a 2.7% improvement in average precision. Additional analysis indicates that the dataset characteristics and the multi-domain fusion strategy contribute to performance enhancement at different stages of the detection pipeline, validating the effectiveness of the proposed dataset for rotated ship detection while demonstrating the complementary role of multi-domain feature enhancement. Full article

Worldwide, breast cancer is the leading cause of death in women. This emphasizes the significance of an accurate breast cancer detection system. This study presents a unified framework for segmentation of breast cancer using multimodal imaging, such as histopathology, MRI, mammogram, and ultrasound. This framework integrates the CNN with Transformer modules and has three core technical innovations. First, features are extracted using an encoder–decoder design. The encoder has Residual Blocks with a base channel of 32, following feature extraction, which are progressively mapped and downsampled into four stages (32 → 64 → 128 → 256) of channels. The spatial channel is reduced using MaxPool2d operations from 256 × 256 to 128 × 128, 64 × 64, 32 × 32, and 16 × 16. After further convolutional refinement, a Transformer encoder is used on the 16 × 16 feature maps in the bottleneck. The Transformer comprises four encoders with multi-head self-attention (eight heads) and a 4.0 MLP ratio, enabling the model to capture local and global contextual dependencies at the lowest resolution. The proposed framework is trained with a learning rate of 1 × 10⁻⁴, up to 50 epochs with early stopping (patience = 12), using a combined Dice and binary cross-entropy loss that balances pixel-wise accuracy and overlap-based learning. Gradient clipping with a maximum norm of 5.0 is used to ensure training stability; ReduceLROnPlateau (factor = 0.5, patience = 5) is used to dynamically adjust the learning rate; and early stopping is used to prevent overfitting. To improve generalization and enhance robustness to data variability, data augmentation techniques such as random horizontal and vertical flips, intensity variations, and small rotations (±15°) are applied. Incremental learning was implemented in this study as a warm-start fine-tuning strategy, where the model was initialized based on learned weights from a previously trained model instead of training from scratch. This is done by loading saved checkpoints of the best-performing model and continuing training on a new dataset. The performance of the proposed framework is evaluated on four publicly available datasets and one local dataset, such as BUS-UCLM, BUSI, BreastDM, TNBC NucleiSegmentation, and BCSD-2024. The impressive results are achieved with Dice scores of 0.974 on ULCM, 0.975 on BUSI, 0.971 on BreastDM, 0.904 on TNBC nuclei segmentation, and 0.982 on BCSD-2024. The proposed model consistently performed better than classical U-Net models. Full article

In remote sensing image oriented object detection tasks, weakly supervised learning methods based on horizontal bounding boxes have attracted much attention due to their lower annotation costs compared to fully supervised methods. However, remote sensing images, characterized by complex backgrounds, exhibit a wide range of target scales and diverse geometric characteristics across target categories. Existing methods exhibit inadequate exploitation of background and angular information under weak supervision, resulting in compromised perception of dense and high-aspect-ratio targets. Neglecting the imbalance in angle estimation samples further leads to excessively low detection accuracy for few-shot categories. To address the aforementioned issues, this paper proposes a Geometry-Aware Enhancement Network (WSOOD-GAEN) for weakly supervised oriented object detection tasks. First, in the backbone network stage, a channel-space deformable attention module (DAE-ResNet) was constructed. Through deformable sampling and screening of key regions, feature extraction has both morphological adaptability to complex shapes and semantic discriminability of key features in complex backgrounds. Secondly, in the feature pyramid stage, an Angle-Guided Feature Pyramid Network (AG-FPN) is proposed. This module dynamically applies rotation transformation to the sampling offsets of deformable convolutions, thereby enhancing the feature representation of objects with different orientations and scales. Furthermore, an adaptive geometric perception loss (AGL) was designed. Based on the geometric characteristics of different categories, it automatically learns differentiated rotation and flip consistency weights, thereby improving the prediction accuracy of small sample categories. Experiments on the DOTA-v1.0, HRSC, and RSAR datasets validate our approach. Specifically, under the AP75 evaluation metric, the proposed method outperforms existing weakly supervised methods by 1.51%, 9.86%, and 3.28%, respectively. Full article

Motor vehicle encroachment into non-motorized lanes is a common but hard-to-verify violation in urban intersections, especially when monitored from unmanned aerial vehicles (UAVs) or high-mounted overhead views. Existing rule-based solutions built on horizontal bounding boxes and center-point/line-crossing criteria are sensitive to perspective distortion, occlusion, and frame-to-frame jitter, resulting in unstable decisions and low evidential value. This paper presents a cascaded UAV-view system that closes the loop from perception to evidence output through detection–segmentation–recognition–decision. First, we adopt a two-stage detection cascade: a lightweight vehicle detector localizes vehicles using axis-aligned bounding boxes, and a dedicated YOLOv5n-based oriented bounding box (OBB) license plate detector, constructed via architecture grafting and weight transfer, is then applied within each vehicle region of interest (ROI) to localize rotated license plates under large pose variation and small-target conditions. Second, a U-Net lane region segmentation module provides pixel-level spatial constraints to define an enforceable lane occupancy region. Third, a perspective rectification step is integrated with the PP-OCRv4 optical character recognition (OCR) framework to improve license plate recognition reliability for tilted plates. Finally, an area ratio criterion and an N-frame temporal counter are used to suppress transient misdetections and stabilize alarms. On a representative 100-sample controlled encroachment benchmark, the proposed system improves detection accuracy from 67.0% to 92.0% and reduces the false positive rate from 32.35% to 5.88% compared with a baseline horizontal bounding box (HBB)-based rule. The system outputs both violation alarms and license plate evidence, supporting practical deployment for multi-view traffic governance. Full article

Withcontinuous advancements in remote sensing technology, object detection in remote sensing images has emerged as a critical research direction in maritime surveillance, port management, and national defense. Among these applications, ship detection is a key task. Due to the fact that ships in images typically exhibit arbitrary rotations, multi-scale distributions, and complex backgrounds, conventional detection methods based on horizontal or rotated bounding boxes often fail to adequately capture the fine-grained information of the targets, thereby compromising detection accuracy. This paper proposes the Dense Representative Points-Guided Rotated-Ship Detection (DenseRRSD) method. The proposed approach represents ship objects using dense representative points (RepPoints) to effectively capture local semantic information, thereby avoiding the background noise issues associated with traditional rectangular bounding box representations. To further enhance detection accuracy, an edge region sampling strategy is devised to uniformly sample RepPoints from critical ship parts, and a Weighted Residual Feature Pyramid Network (WRFPN) is introduced to efficiently fuse the multi-scale features through residual connections and learnable weights. In addition, a Weighted Chamfer Loss (WCL) combined with a staged localization loss strategy is employed to progressively refine localization from coarse to fine stages. Experimental results on both the HRSC2016 dataset and the newly constructed DOTA-SHIP dataset demonstrate that DenseRRSD achieves state-of-the-art detection accuracy, with mean Average Precision (mAP) scores of 91.2% and 83.2%, respectively, significantly outperforming existing methods. These results verify the effectiveness and robustness of the proposed approach in rotated-ship detection under diverse conditions. Full article

Accurate measurement of sea-surface winds is critical for climate science, physical oceanography, and the rapidly expanding offshore wind energy sector. Buoy-based platforms—moored meteorological buoys, drifters, and floating LiDAR systems (FLS)—provide practical alternatives to fixed offshore structures, especially in deep water where bottom-founded installations are economically prohibitive. Yet these floating platforms are subject to continuous pitch, roll, heave, and yaw motions forced by wind, waves, and currents. Such six-degree-of-freedom dynamics introduce multiple error pathways into the measured wind signal. This paper synthesizes the current understanding of motion-induced measurement errors and the techniques developed to compensate for them. We identify four principal error mechanisms: (1) geometric biases caused by sensor tilt, which can underestimate horizontal wind speed by 0.4–3.4% depending on inclination angle; (2) contamination of the measured signal by platform translational and rotational velocities; (3) artificial inflation of turbulence intensity by 15–50% due to spectral overlap between wave-frequency buoy motions and atmospheric turbulence; and (4) beam misalignment and range-gate distortion specific to scanning LiDAR systems. Compensation strategies have progressed through four recognizable stages: fundamental coordinate-transformation and velocity-subtraction algorithms developed in the 1990s; Kalman-filter-based multi-sensor fusion emerging in the 2000s; Response Amplitude Operator modeling tailored to FLS platforms in the 2010s; and data-driven machine-learning approaches under active development today. Despite this progress, key challenges persist. Sensor reliability degrades under extreme sea states precisely when accurate data are most needed. The coupling between high-frequency platform vibrations and turbulence remains poorly characterized. No unified validation framework or benchmark dataset yet exists to compare methods across platforms and environments. We conclude by outlining research priorities: end-to-end deep-learning architectures for nonlinear error correction, adaptive algorithms capable of all-sea-state operation, standardized evaluation protocols with open datasets, and tighter integration of intelligent software with next-generation low-power sensors and actively stabilized platforms. Full article

A combined numerical simulation and experimental validation approach was employed to investigate the phenomena of screw adhesion and nozzle clogging, which occur frequently during material conveying and extrusion of high-viscosity, high-solid-content multi-component energetic materials in additive manufacturing. First, conical and cylindrical screws were designed. Through simulation calculations of the energetic material extrusion process, patterns in the variation in internal pressure and shear rate within the screw were analyzed, providing guidance for the design of the printing equipment. Second, a Z-shaped stirring paddle kneading device and a dual-nozzle printing device featuring horizontally and vertically arranged two-stage screws were designed. Through extrusion experiments with PBX (polymer-bonded explosive) slurry, the optimal matching relationship between the kneading rate and the extrusion rates of the horizontal and vertical screws was obtained. Finally, additive manufacturing of complex-shaped PBX charges using high-viscosity energetic materials was successfully accomplished. This confirms the further optimization of the additive manufacturing equipment in terms of safety control, precision control, and adaptability to complex structures under extreme operating conditions. The results indicate that the cylindrical screw outperforms the conical screw, and with a screw clearance of 3mm, it represents the optimal design solution. During the kneading process, a screw rotational speed of 25 rpm was used. After kneading for 3 h, the slurry exhibited good uniformity, with a solid content of approximately 70% and relatively small deviation. During the extrusion process, a nozzle diameter of 1.55 mm combined with a rotational speed of 5 rpm for the horizontal screw (feeding screw) and 7 rpm for the vertical screw (extrusion screw) can satisfy the requirements of the “starved feeding” mode, thus achieving continuous and stable filament formation of the slurry. Full article

This study presents the design and development of an intelligent end-effector integrated into a custom 7-degree-of-freedom (DOF) robotic arm for monitoring the health status of tomato plants during their growth stages. The robotic system combines five rotational and two prismatic joints, enabling both horizontal reach and vertical adaptability to inspect plants of varying heights without repositioning the robot’s base. The integrated vision module employs a YOLOv5 neural network trained with 7864 images of tomato leaves, including both healthy and diseased samples. Image preprocessing included normalization and data augmentation to enhance robustness under natural lighting conditions. The optimized model achieved a detection accuracy of 90.2% and a mean average precision (mAP) of 92.3%, demonstrating high reliability in real-time disease classification. The end-effector, fabricated using additive manufacturing, incorporates a Raspberry Pi 4 for onboard processing, allowing autonomous operation in agricultural environments. The experimental results validate the feasibility of combining a custom 7-DOF robotic structure with a deep learning-based detector for continuous plant monitoring. This research contributes to the field of agricultural robotics by providing a flexible and precise platform capable of early disease detection in dynamic cultivation conditions, promoting sustainable and data-driven crop management. Full article

Background: The early and accurate classification of ischemic stroke stages on computed tomography (CT) remains challenging due to subtle attenuation differences and significant scanner variability. This study developed a neural network framework to dynamically optimize Hounsfield Unit (HU) transformations and CLAHE parameters for temporal stage-specific stroke classification. Methods: We analyzed 1480 CT cases from 68 patients across five stages (hyperacute, acute, subacute, chronic, and normal). The training data were augmented via horizontal flipping, ±7° rotation. A convolutional neural network (CNN) was used to optimize linear transformation and CLAHE parameters through a combined loss function incorporating the effective measure of enhancement (EME), peak signal-to-noise ratio (PSNR), and regularization. the enhanced images were classified using logistic regression (LR), support vector machines (SVMs), and random forests (RFs) with 25-fold cross-validation. Model interpretability was evaluated using Grad-CAM. Results: Neural network optimization significantly outperformed static parameters across validation metrics. Deep CLAHE achieved the following accuracies versus static CLAHE: hyperacute (0.9838 vs. 0.9754), acute (0.9904 vs. 0.9873), subacute (0.9948 vs. 0.9825), and chronic (near-perfect 0.9979 vs. 0.9808). Qualitative interpretability analysis confirmed that models focused on clinically relevant regions, with optimized enhancement producing more coherent attention patterns than static methods. Parameter analysis revealed stage-aware adaptation: conservative enhancement in early phases (slope: 1.249–1.257), maximized in subacute (slope: 1.290–1.292), and restrained in the chronic phase (slope: 1.240–1.258), reflecting underlying stroke pathophysiology. Conclusions: A neural network-optimized framework with interpretability validation provides stage-specific stroke classification that achieves superior performance over static methods. Its pathophysiology-aligned parameter adaptation offers a clinically viable and transparent solution for emergency stroke assessment. Full article

The increasing burial depth of deep shale formations in the southern Sichuan leads to more complex in situ stresses and geological structures, which in turn raises the challenges of hydraulic fracturing. Although enlarging the treatment scale and injection rate can enhance reservoir stimulation, the intensive development of faults and fractures in deep shale formations aggravates stress instability, inducing casing deformation, fracture communication, and other engineering risks that constrain efficient shale gas production. In this study, a cross-scale geomechanical model linking the regional to near-wellbore domains was constructed. A dynamic evolution equation was established based on flow–stress coupling, and a numerical conversion from the geological model to the finite element model was implemented through self-programming, thereby developing a simulation method for dynamic geomechanical evolution during hydraulic fracturing. Results indicate that dynamic variations in pore pressure dominate stress redistribution, while near-wellbore heterogeneity and mechanical property distribution significantly affect prediction accuracy. The injection of fracturing fluid generates a high-pressure gradient that drives pore pressure diffusion along fracture networks and faults, exhibiting a strong near-wellbore but weak far-field non-steady spatial attenuation. As the pore pressure increases, the peak value reaches 1.4 times the original pressure. The triaxial stress shows a negative correlation and decreases. The horizontal minimum principal stress shows the most significant drop, with a reduction of 15.79% to 20.68%, while the vertical stress changes the least, with a reduction of 5.7% to 7.14%. Compared with the initial stress state, the horizontal stress difference increases during fracturing. Rapid pore-pressure surges and fault distributions further trigger stress reorientation, with the magnitude of rotation positively correlated with the intensity of pore-pressure variation. The high porosity and permeability characteristics of the initial fracture network lead to a rapid attenuation of the stress around the wellbore. In the middle and later stages, as the pressure balance is achieved through fracture filling, the pore pressure rises and the stress decline tends to stabilize. The findings provide significant insights into the dynamic stress evolution of deep shale reservoirs during fracturing and offer theoretical support for optimizing fracturing design and mitigating engineering risks. Full article

Background/Objectives: Artificial intelligence (AI) has been increasingly used to support radiological assessment in dentistry. We benchmarked machine learning (ML), deep learning (DL), and a hybrid optimization-assisted approach for the automatic five-class image-level classification of dental restorations (filling, implant, root canal treatment, fixed partial denture/bridge, crown) on panoramic radiographs. Methods: We analyzed 353 anonymized panoramic images comprising 2137 labeled restorations, acquired on the same device. Images were cropped and enhanced (histogram equalization and CLAHE), and texture features were extracted with GLCM. A three-stage pipeline was evaluated: (i) GLCM-based features classified by conventional ML and a baseline DL model; (ii) Hybrid Grey Wolf–Particle Swarm Optimization (HGWO-PSO) for feature selection followed by SVM; and (iii) a CNN trained end-to-end on raw images. Performance was assessed with an 80/20 per-patient split and 5-fold cross-validation on the training set. While each panoramic radiograph may contain multiple restorations, in this study we modeled the task as single-label, image-level classification (dominant restoration type) due to pipeline constraints; this choice is discussed as a limitation and motivates multi-label, localization-based approaches in future work. The CNN baseline was implemented in TensorFlow 2.12 (CUDA 11.8/cuDNN 8.9) and trained with Adam (learning rate 1 × 10⁻⁴), with a batch size 32 and up to 50 epochs with early stopping (patience 5); data augmentation included horizontal flips, ±10° rotations, and ±15% brightness variation. A post hoc power analysis (G*Power 3.1; α = 0.05, β = 0.2) confirmed sufficient sample size (n = 353, power > 0.84). Results: The HGWO-PSO + SVM configuration achieved the highest accuracy (73.15%), with macro-precision/recall/F1 = 0.728, outperforming the CNN (68.52% accuracy) and traditional ML models (SVM 67.89%; DT 59.09%; RF 58.33%; K-NN 53.70%). Conclusions: On this single-center dataset, the hybrid optimization-assisted classifier moderately improved detection performance over the baseline CNN and conventional ML. Given the dataset size and class imbalance, the proposed system should be interpreted as a decision-supportive tool to assist dentists rather than a stand-alone diagnostic system. Future work will target larger, multi-center datasets and stronger DL baselines to enhance generalizability and clinical utility. Full article