Search Results (577)

Medical image classification in clinical settings is frequently constrained by High-Dimensional, Low-Sample Size (HDLSS) conditions, rendering conventional Support Vector Machines (SVM) geometrically susceptible to the data piling phenomenon. This study proposes a hybrid CNN-DWD framework to address this geometrical instability by integrating multi-architecture convolutional feature extraction with Distance-Weighted Discrimination (DWD). Pre-trained ResNet50 and DenseNet121 backbones act as frozen feature extractors, generating a highly descriptive 3072-dimensional fused representation. To resolve the computational bottleneck of deploying DWD directly on massive feature spaces, a supervised Multi-Layer Perceptron (MLP) bottleneck progressively compresses this space into a 32-dimensional latent manifold. Evaluated across breast ultrasonography, breast mammography, and chest X-ray datasets under varying training allocations, the proposed architecture drastically accelerates DWD training—achieving over a 130-fold speedup. The proposed CNN-MLP-DWD framework demonstrates highly competitive diagnostic performance, achieving 93.16% accuracy on the breast ultrasonography dataset and a macro-AUC of 99.69% on the chest X-ray benchmark, comparing favorably against the evaluated baseline methods. Full article

Accurate and interpretable brain tumour detection from magnetic resonance imaging (MRI) is important for the reliable use of computer-assisted diagnostic systems. This study examines whether image preprocessing can improve the localisation quality of explanations generated by convolutional neural network (CNN) classifiers while preserving high classification performance. Two pre-trained CNN architectures, ResNet50 and DenseNet121, were fine-tuned using the BRISC 2025 dataset, which contains 6000 annotated contrast-enhanced T1-weighted MRI images: 5000 training images and 1000 test images. The dataset includes four classes: glioma, meningioma, pituitary tumour, and healthy brain images. The original classification layers were replaced with custom fully connected heads designed for four-class classification. Model explanations were generated using Grad-CAM, Integrated Gradients, and LIME. Their localisation quality was evaluated against the available tumour segmentation masks using Intersection over Union (IoU), the Dice coefficient, and the Pointing Game metric. Tests show that both models (99.1% for ResNet50 and 99.2% for DenseNet121) perform well in terms of validation accuracy, but the explanation maps often operate in regions outside the clinically relevant area. To combat this issue, an image preprocessing pipeline utilising Otsu threshold masking, hole filling, and brightness–contrast jittering was implemented to filter noise from the background and isolate the focus area—brain region. After preprocessing, the validation accuracy of the DenseNet121 model was 99.6%, and the average Grad-CAM explainability metrics improved from 21.49% to 23.08% IoU, from 30.21% to 32.49% Dice coefficient, and from 48.33% to 51.67% Pointing Game score. The results indicate that conventional image preprocessing can moderately improve the spatial agreement between explanation maps and annotated tumour regions without reducing classification accuracy. Full article

Background: Cervical cancer remains one of the leading causes of cancer-related morbidity among women worldwide. The Papanicolaou (Pap) smear is widely used for early detection; however, its manual interpretation is time-consuming, requires substantial expertise, and is often affected by inter-observer variability, particularly in cases with dense and overlapping cells. Methods: This study developed an optimized deep learning framework for cervical cell instance segmentation, specifically targeting the separation of overlapping cells in Pap smear images. The proposed framework was based on Mask R-CNN with a ResNet-50 backbone and Feature Pyramid Network. A public development dataset of 460 cervical smear images was used for model development and internal evaluation, while an independent 210-image dataset collected from King Abdullah University Hospital was reserved as an external held-out clinical assessment set. To improve small-cell detection and mask refinement, the Region Proposal Network was adapted using biologically informed anchor scales (16, 32, 64, 128, 256), followed by soft calibration-based post-processing. Results: The proposed framework achieved an AP50 of 89.90% and a Mask IoU of 90.44%. Small-cell performance, measured using APs under the COCO small-object convention, reached 22.80%, while APm and APl reached 68.45% and 81.42%, respectively. Clinical cell-count evaluation on the independent KAUH held-out set was performed using the mean count of five physicians as the reference standard and achieved an overall clinical detection accuracy of 93.00%. Conclusions: The optimized Mask R-CNN framework improved the detection and separation of overlapping cervical cells in Pap smear images and may serve as a supportive tool for cytopathology workflows. The results suggest that biologically informed anchor optimization and soft calibration can improve cell-level instance segmentation, particularly for small and overlapping cells. Further validation on larger multi-center datasets remains necessary before routine clinical deployment. Full article

Background/Objectives: Colorectal cancer (CRC) is the second leading cause of cancer-related mortality worldwide, with polyp miss rates of up to 26% reported during colonoscopy and classification accuracy remaining highly operator-dependent. Accurate multi-class polyp subtype classification is clinically critical, as it directly determines treatment decisions: adenomatous polyps require resection, whereas hyperplastic lesions may warrant only surveillance. This study aims to systematically compare five deep learning architectures for five-class gastrointestinal polyp classification and to provide clinically interpretable diagnostic insights through Grad-CAM visualisation. Methods: ResNet50, VGG16, EfficientNet-B3, DenseNet121, and Vision Transformer (ViT-B/16) were evaluated on the Kvasir Dataset V2 (5000 images, five classes) under a unified training and evaluation protocol on common GPU hardware. All models employed ImageNet transfer learning with a redesigned multi-layer classification head. Two optimisation strategies were applied: SGD with cosine annealing for CNN architectures, and AdamW with linear warmup for ViT-B/16. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to generate spatial attention heatmaps for qualitative clinical interpretation. Results: Under a single 80/10/10 split, ViT-B/16 attained the highest accuracy (97.2%); however, because a single split is sensitive to sampling, the evaluation was strengthened with stratified five-fold cross-validation (mean ± SD). Under cross-validation, EfficientNet-B3 achieved the highest accuracy at 95.90 ± 0.35%, followed closely by ViT-B/16 (95.12 ± 0.72%), then ResNet50 (91.74 ± 0.74%), DenseNet121 (90.32 ± 0.70%), and VGG16 (88.50 ± 1.72%); the small standard deviations indicate that all models, including ViT-B/16, were stable across folds. Pairwise McNemar tests with Holm correction found that every difference was statistically significant (p < 0.05), including the EfficientNet-B3 advantage over ViT-B/16 (p = 0.010). ViT-B/16 thus remained a strong, stable performer that significantly outperformed the three remaining CNNs, while the cross-validated ranking placed the most compact model, EfficientNet-B3, first: a Vision Transformer was highly competitive with, but not superior to, the strongest CNN. A consistent, architecture-agnostic misclassification pattern was identified between dyed-lifted polyps and dyed-resection margins across all five models, consistent with a task-level visual ambiguity that may also reflect overlapping class definitions and annotation factors, with direct clinical implications. Grad-CAM analysis, quantified by attention-entropy and concentration metrics, showed that model attention remained focused on relevant stained tissue regardless of whether predictions were correct, indicating that the dyed-class confusions reflect genuine visual ambiguity rather than a localisation failure. Conclusions: Under cross-validation, EfficientNet-B3 achieved the highest accuracy on the Kvasir V2 five-class task, significantly outperforming all other architectures, with ViT-B/16 being a close and competitive second. The identified confusion between post-procedural chromoendoscopic classes is unlikely to be fully resolved by architectural changes alone and may require higher-resolution imaging or domain expert re-annotation. These findings contribute to the evidence base for explainable deep learning in gastrointestinal endoscopy; external, multi-centre validation remains necessary before clinical adoption. Full article

Background/Objective: The management and prevention of impacted maxillary canines is both an art and a science. These important teeth serve both aesthetic and functional purposes, so they demand meticulous evaluation. Deep learning model tools are game-changing technologies, elevating diagnostic capabilities and therapeutic planning. This study comprised a comparative analysis of the prediction of permanent maxillary canine eruption as a preventive measure between experts and a deep learning model. Methods: In this retrospective cross-sectional study, 2230 panoramic radiographs of patients aged 9–14 years were analyzed to assess the patterns of eruption of unerupted maxillary permanent canines (UPMCs). The study images were classified into three sectors according to the modified Ericson and Kurol sectors. Data preprocessing techniques were used to prepare images for the deep learning model by using a DenseNet121-based Convolutional Neural Network (CNN). The data were split into training and testing sets to train the AI to predict sectors. The deep learning model’s predictions were evaluated using sector accuracy, precision, recall, and F1 score. Results: The sample included 796 (35.6%) men and 1434 (64.3%) women. The mean age of the participants was 12.2 ± 1.60 years. For sector 1, the right side of UMPCs, the AI reported an accuracy of 99.55% and perfect precision at 100%, while for the left side, accuracy and precision were 99.24% and 100.00%, respectively. For sector 2, prediction on the right side, the performance accuracy values reached 99.73%, and the precision was 98.15%. For the left side of UMPCs, the prediction accuracy was 98.79%, and precision was 94.10%. Regarding sector 3, the right side saw 99.82% accuracy and 98.83% precision. For the left side, the AI achieved an overall accuracy of 99.46% and precision of 99.72%. Conclusions: The deep learning-based system significantly reduced the time and human resources required for landmark identification and parameter generation, making the diagnostic process more efficient in interceptive/preventive orthodontics. Full article

Pediatric wrist fractures are among the most prevalent musculoskeletal injuries in children. Fracture subtype, including buckle/torus, greenstick, and Salter–Harris physeal injuries, directly influences management and prognosis. Subspecialty radiographic expertise required for subtype classification is not universally available in emergency or resource-limited settings. Deep learning (DL) offers an automated approach to fracture subtype recognition from plain radiographs. This pilot study evaluated convolutional neural network (CNN)-based five-class pediatric wrist fracture classification using the GRAZPEDWRI-DX dataset.A total of 940 pediatric wrist radiographs from GRAZPEDWRI-DX (figshare ID 14825193) were labeled using Arbeitsgemeinschaft fur Osteosynthesefragen (AO) pediatric codes into five classes: no fracture, buckle/torus, greenstick, Salter–Harris physeal fracture, and other fracture. Contrast-limited adaptive histogram equalization (CLAHE) and letterbox resizing to 224 × 224 pixels were applied. Patient-level stratified splits (70/15/15%) prevented data leakage. Three ImageNet-pretrained architectures (DenseNet-169, ResNet-50, and EfficientNet-B4) underwent two-phase transfer learning. Performance was assessed by balanced accuracy, macro F1, macro area under the receiver operating characteristic curve (AUROC), and Cohen’s kappa.DenseNet-169 achieved the highest balanced accuracy (0.371; 95% confidence interval [CI]: 0.289–0.448), macro F1 (0.334; 95% CI: 0.251–0.416), and macro AUROC (0.669), with Cohen’s kappa of 0.269 on the held-out test set (n = 139) under initial five-epoch pilot training conditions. All three networks exceeded a majority-class (no-information) baseline (balanced accuracy 0.20). Extending training to 50 epochs (approximately 2100 mini-batch iterations) with GPU acceleration substantially improved DenseNet-169 to a balanced accuracy of 0.532 (95% CI: 0.451–0.614), macro F1 of 0.516, and macro AUROC of 0.815, with statistically significant pairwise architecture differences (McNemar p < 0.01); per-class sensitivity was highest for no-fracture detection (0.969) and lowest for buckle/torus fractures (0.393). Gradient-weighted class activation mapping (Grad-CAM) confirmed anatomically coherent model saliency at the distal radial metaphysis and physeal plate.DenseNet-169 achieved the best five-class classification performance among evaluated architectures under pilot training conditions, and extended training substantially improved accuracy, although classification accuracy remained below clinically usable thresholds. These results establish a reproducible, patient-stratified DL pipeline and a benchmark for full-dataset training and future methodological development, rather than a clinically deployable tool. Full article

The BQ method, the most widely used technique for classifying tunnel surrounding rocks, requires correction factors including groundwater conditions, structural surface characteristics, and initial ground stress. However, existing deep learning approaches address only a single parameter and rely heavily on empirical judgment, lacking the capability for comprehensive multi-parameter intelligent assessment. To address these limitations, a large-scale database of 21,800 tunnel face images was constructed through on-site data collection, data aggregation, and image augmentation. Five convolutional neural network (CNN) models were trained and evaluated using a proposed multi-indicator scoring method comprising seven performance metrics: loss value, accuracy, precision, recall, confusion matrix, frames per second (FPS), and model size. EfficientNet-B2, ResNet101, and DenseNet121 achieved the highest scores of 100, 84, and 87 for groundwater classification, rock structure type classification, and weathering degree classification, respectively. These three optimized models were integrated into a unified software platform that recognizes multiple geological attributes from a single tunnel face image. Field validation across multiple sections of the Jiaozhou Bay Second Undersea Tunnel shows that the platform achieves a recognition accuracy of over 90%. The results demonstrate that the proposed multi-indicator evaluation method yields more-comprehensive model selection, and the integrated platform can directly support BQ value correction, contributing to intelligent surrounding rock classification in complex tunnel construction environments. Full article

Background/Objectives: Osteoporosis is a major public health concern associated with increased fracture risk and reduced quality of life if not detected at an early stage. Automated analysis of knee X-ray images using artificial intelligence has shown promising potential for opportunistic osteoporosis screening. This study aims to develop and evaluate a multi-stage deep learning and machine learning framework for osteoporosis classification, with particular emphasis on external validation, calibration drift, and cross-domain generalization performance. Methods: Knee X-ray images were categorized into three classes: Normal, Osteopenia, and Osteoporosis. Deep features were extracted using pretrained convolutional neural networks, including ResNet18, EfficientNetB0, and DenseNet121. The extracted features were subsequently classified using multiple machine learning models, including Neural Network, Efficient Linear, Support Vector Machine, and Naive Bayes classifiers. Two data augmentation strategies were investigated: targeted minority-class augmentation and full 3× dataset expansion with class balancing. Model performance was evaluated using accuracy, precision, recall, F1-score, and AUC on internal validation, independent test sets, and external validation datasets. Additional analyses included reliability calibration assessment, isotonic recalibration, and class-prior boosting with cross-validated threshold optimization to address external domain shift. Results: EfficientNetB0 and DenseNet121 consistently outperformed ResNet18 across most evaluation metrics. Under the balanced augmentation strategy, EfficientNetB0 combined with Efficient Linear demonstrated strong and stable performance, while DenseNet121 paired with a Neural Network achieved the highest overall classification performance. External validation revealed a substantial discrepancy between AUC and threshold-based metrics, indicating the presence of calibration drift and class-prior mismatch across imaging domains. Reliability analysis showed severe probability collapse in the Osteopenia class during external testing. Post-hoc recalibration improved probability reliability, while class-prior boosting substantially increased Osteopenia sensitivity and improved balanced accuracy and macro F1-score under external validation conditions. Conclusions: The proposed framework demonstrates the feasibility of combining pretrained CNN-based deep feature extraction with machine learning classifiers for osteoporosis classification from knee X-ray images. The findings further highlight that maintaining model performance under external testing conditions may require not only strong feature extraction capability but also adaptive recalibration and deployment-aware threshold optimization to address calibration drift and cross-domain variability. While the results are encouraging, the present study should be considered a proof-of-concept investigation. Although the framework was evaluated using an independent public external dataset, further validation using larger and more diverse multi-center clinical cohorts is necessary to establish generalizability and clinical utility before routine clinical implementation can be considered. Full article

Background: Malaria is a life-threatening disease caused by Plasmodium species, which is endemic in tropical and subtropical regions worldwide. In clinical settings, experienced parasitologists perform microscopic examinations of thick/thin blood slides. This method is labour-intensive and is adversely affected by inter- and intra-observer variability among the microscopists. The present study aimed to develop a malaria screening algorithm using computer vision to identify and classify malaria parasite-infected red blood cells (RBC) from microscopic blood slide images. Methods: The proposed classification methodology first employs digital image processing techniques, the watershed transform, to preprocess the raw images, followed by connected component labelling to accurately segment and isolate individual RBCs from the background. To classify these segmented cells as either normal or infected, convolutional neural networks (CNNs) were utilized, leveraging their ability to automatically extract relevant features through deep, hidden layers, thus eliminating the need for manual feature engineering. Results: To compare and determine the most effective classification engine, the study developed and evaluated five distinct models: four well-established transfer learning architectures (VGG16, VGG19, DenseNet121, and InceptionV3), alongside a newly proposed custom CNN model. A total of 2422 segmented RBC images were used for the training, and 692 different images were used for testing, with the VGG model showing the best accuracy at 99.57%. The proposed CNN architecture also showed competitive results with 99.14% accuracy. Conclusions: Transfer learning models demonstrated remarkable accuracy for malaria parasite classification from blood smear slides, with VGG19 (99.57%) achieving the highest accuracy on diverged datasets for the test images. The analysis demonstrates the potential of this approach as a computational aid for future image-based malaria screening in conjunction with existing diagnostic tests. Full article

The welded bead bending test (WBBT) serves as a pivotal procedure for evaluating the crack-arrest capacity of structural steels used for construction of safety-critical infrastructure according to ZTV-ING Part 4 or Deutsche Bahn Standard 918 002-02. Previous research has established that light optical microscopy images (LOMs) of specimens that do not pass the WBBT are frequently characterised by a high prevalence of Widmanstätten ferrite within the base material. While convolutional neural networks (CNNs) have successfully classified WBBT outcomes based on LOMs, the implemented approaches did not enable simultaneous pixel-wise delineation required for accurate quantification of Widmanstätten ferrite. However, the tonal similarity between Widmanstätten ferrite and polygonal ferrite renders conventional intensity-based thresholding ineffective for differentiation, necessitating a morphology-based segmentation approach. The present study proposes an automated semantic segmentation framework for the precise delineation of three microstructural phases present within the WBBT LOMs: polygonal ferrite, pearlite, and Widmanstätten ferrite. Utilising a dataset of 20 LOMs and corresponding manually annotated masks, five UNet encoder backbones were evaluated: ResNet-152, Xception, SE-ResNet-50, DenseNet-169, and EfficientNet-B5. To identify the optimal configuration for this microstructural segmentation task, each architecture was assessed using three distinct weight initialisation strategies: (A) ImageNet, (B) MicroNet, and (C) a combined approach. For Widmanstätten ferrite segmentation at patch level, DenseNet-169 pretrained on ImageNet achieves the best performance (Dice: 50.75% ± 3.38%, IoU: 38.45% ± 3.01%). Following inference-based aggregation, Xception pretrained on ImageNet yields improved results (Dice: 68.74% ± 1.12%, IoU: 52.52% ± 1.28%), with an MAE of 4.55% ± 0.80%. Full article

Post-harvest losses of mango (Mangifera indica L.) in developing economies are estimated at 5% to 30%, largely due to manual management practices that depend on subjective visual assessments. This paper proposes a lightweight deep learning architecture, termed SE-DBIRNet, for real-time classification of mangoes into five shelf-life stages: unripe, semi-ripe, fully ripe, overripe, and perished. The model incorporates three key design strategies: (i) depthwise separable convolutions, achieving an 88.5% reduction in parameters when integrated into the ResNet50 backbone; (ii) a double-branch inverted residual (DBIR) module designed to enhance feature diversity and richness; and (iii) a squeeze-and-excitation (SE) attention mechanism for adaptive channel-wise recalibration. Using a public benchmark dataset of 4428 RGB images (Mendeley Data) under 10-fold cross-validation, SE-DBIRNet achieved 98.24% accuracy. Among lightweight CNN architectures (EfficientNetB0, MobileNetV2, ResNet50), SE-DBIRNet outperformed the best lightweight baseline (EfficientNetB0: 96.57%) by 1.67 percentage points. While dedicated attention-based DenseNet variants (e.g., DSA-DenseNet: 99.20%) achieved higher accuracy, SE-DBIRNet offers a superior trade-off among accuracy, inference speed (56.9 ± 1.8 FPS), and memory efficiency (8871 ± 45 MB CPU memory). EigenCAM activation visualizations revealed that the model focuses on biologically relevant and stage-discriminative features, including surface color gradients, texture uniformity, lenticel patterns, and decay boundaries. Overall, SE-DBIRNet achieves a Pareto-optimal balance among accuracy, speed, and memory efficiency, making it a strong candidate for real-time, edge-deployable post-harvest mango quality-monitoring systems, particularly when computational resources are limited. Full article

Glaucoma, the leading cause of irreversible vision loss, often goes undetected in early stages due to its asymptomatic behaviour. Early diagnosis typically involves visual analysis of the optic disc (OD) in eye fundus images. Machine and deep learning techniques have emerged as valuable tools for automating this process; however, their integration into clinical practice still faces limitations. These challenges include the presence of image regions that are not directly related to glaucoma assessment, such as retinal vasculature, the macula, and background structures, which may introduce irrelevant information and negatively affect classification performance, as well as a general lack of transparency in the decision-making process. This article proposes a methodology that enhances both the accuracy and interpretability of glaucoma detection by focusing solely on the OD region. First, a metaheuristic-based strategy is employed for precise OD detection and cropping, generating an OD-centric dataset with glaucoma-labeled images, which is composed of different public datasets. Four convolutional neural networks (CNNs), namely VGG-19, MobileNet-V2, ResNet-50, and DenseNet-161, are trained on this dataset using transfer learning. To address the need for model explainability, Grad-CAM, Score-CAM, and Eigen-CAM are applied to the trained models to generate post hoc visual explanations of their predictions. The experimental results showed that DenseNet-161 achieved the best overall performance on the assembled public dataset, using an 80%-10%-10% training, validation, and testing split, with a test accuracy of 0.9369 and an AUC of 0.9831. By isolating the OD region and incorporating explainability techniques, the methodology provides a robust and interpretable second opinion, supporting more accurate and efficient glaucoma screening. Full article

The early and accurate detection of lung cancer (LC) is one of the primary challenges in the clinical diagnostics process, which plays a vital role in the treatment of the disease. Although various deep learning (DL) techniques have been presented, the existing DL methods are mainly focused on single-modal images, either computed tomography (CT) or histopathological images, which are associated with poor generalization, diversity, and applicability. To mitigate the existing issues, the present work aims to develop a modality-independent ensemble DL framework that is independently evaluated on CT and histopathological image datasets for LC classification. In this work, the proposed framework was developed using the Beta Normalization Aggregation (BNA) technique, where the performance of three state-of-the-art pre-trained convolutional neural network (CNN) architectures was compared on two distinct imaging modalities images. Based on the comparative analysis of the performance metrics, Xception, DenseNet121, and MobileNetV2, are chosen to develop the Ensemble model. Predictions generated by the selected CNN models are aggregated using the proposed BNA strategy to improve classification robustness, which improves the confidence of the prediction results and discriminative capabilities. The experiments using public data sets have confirmed the excellent performance of the model. On the CT dataset, the proposed BNA Ensemble achieved a testing accuracy of 97.45%, with a precision of 97.88%, recall of 97.45%, F1-score of 97.45%, and an AUC of 0.9986. On the histopathological dataset, the framework achieved an accuracy of 99.80%, with precision, recall, and F1-score all reaching 99.80%, and an AUC of 1.0000. These results demonstrate the effectiveness, robustness, and generalizability of the proposed BNA framework. The analysis of the results using t-SNE plots, confusion matrices, ROC curves, and confidence distributions provided additional insights into feature separability, classification performance, and prediction confidence of the proposed framework. Full article

Manual classification of rice grain defects remains a persistent challenge in the Peruvian rice industry, as it relies heavily on human inspection, leading to variability, inconsistency, and reduced efficiency when processing large volumes of product. This study evaluates the effectiveness of transfer learning and convolutional neural networks (CNNs) for the automatic classification of four rice grain categories relevant to quality assessment: Whole, Stained, Broken, and Chalky. A dataset comprising 6599 RGB images was employed. To ensure a reliable evaluation protocol, the dataset was first partitioned into training (70%), validation (15%), and test (15%) subsets, after which data augmentation was independently applied within each partition to balance class distributions. Five pretrained CNN architectures were evaluated: MobileNetV2, EfficientNetB0, ResNet50, DenseNet121, and InceptionV3, all of which share a common classification head. Models were trained using transfer learning and early stopping based on validation loss. Performance was assessed using accuracy, precision, recall, F1-score, confusion matrices, 95% confidence intervals, and pairwise McNemar statistical tests. The results showed that ResNet50 achieved the highest classification accuracy (84.71%), followed by EfficientNetB0 (83.60%) and DenseNet121 (83.20%). Statistical analysis indicated that performance differences among the top-performing architectures were relatively small, with significant differences observed only for selected model pairs. Across all evaluated models, the discrimination between Whole and Chalky grains remained the most challenging classification task due to their high visual similarity. Overall, the findings demonstrate that transfer learning-based CNNs provide an effective and scalable approach for automated rice grain defect identification and quality assessment in agricultural environments. Full article

High-resolution remote sensing image segmentation is important for urban mapping but remains challenging because of spectral ambiguity, large scale variations, fragmented elongated structures, and background interference. This study aims to improve semantic segmentation in complex aerial scenes by combining local feature extraction, selective multi-scale fusion, and global sequence modeling. We propose DSD-Mamba, an asymmetric dual-stream architecture with a ResNet-18 encoder. The Dense-Sparse Pyramid Fusion Module aligns multi-level features and applies dual Top-k selective value aggregation for cross-scale response filtering and background-response suppression. This Top-k operation is used as a feature-selection mechanism and is not intended to reduce the theoretical memory footprint of dense attention. Scale-Aware Strip Attention refines skip connections through horizontal and vertical dependency modeling, and the Dual-Stream Context Decoder combines a Mamba-based global branch with a CNN-based local branch during upsampling. Experiments were conducted on UAVid, ISPRS Vaihingen, and ISPRS Potsdam under a single-model inference protocol without test-time augmentation. DSD-Mamba achieved mIoU scores of 73.4%, 85.2%, and 87.2%, respectively. Ablation experiments on Vaihingen showed that DSPFM, SASA, and DSCD improved performance over the baseline when evaluated in this setting, with the full model reaching the highest mIoU. The method improves segmentation accuracy under the tested protocols, although its higher FLOPs indicate an accuracy-oriented rather than lightweight design. Full article