U-Net-Based Model Design for Semantic Segmentation of Class-Imbalanced Semi-Synthetic Roads

Infrastructure cracks are critical indicators of structural deterioration in pavements, bridges, and buildings. Automated crack segmentation has therefore become an important component of structural health monitoring systems. However, accurate pixel-level crack segmentation on resource-constrained devices remains challenging due to the thin, low-contrast, and curvilinear morphology of cracks, as well as severe foreground-background class imbalance. To address these challenges, we propose LiteCrackSeg, a lightweight hybrid CNN-transformer architecture designed for efficient and accurate crack segmentation. The proposed framework adopts a hybrid MobileViT encoder that captures both local spatial details and long-range contextual dependencies while maintaining a compact model size. To enhance morphological sensitivity to elongated crack structures, we introduce a Morphology-Aware MobileViT (MAM-ViT) bottleneck, which integrates dual-branch Dynamic Snake Convolutions (DSConv) to align receptive fields with crack trajectories. Furthermore, a transformer-based decoder with local self-attention progressively reconstructs spatial details, while an attention-guided multi-scale fusion strategy improves boundary precision and structural continuity. To mitigate severe class imbalance, the model is trained using the Tversky loss, which explicitly balances false positives and false negatives. Extensive experiments on three public crack segmentation datasets (DeepCrack, CrackMap, and TUT) demonstrate that LiteCrackSeg achieves state-of-the-art segmentation performance while maintaining high computational efficiency. The proposed model requires only 2.72M parameters and 3.23 GFLOPs, achieving real-time inference at 56 FPS on 512 × 512 images, making it suitable for deployment on resource-constrained edge devices for practical infrastructure inspection.

Class imbalance and the curse of minority hubs

Optical coherence tomography plays a critical role in diagnosing retinal diseases, yet automated deep learning classification is hindered by severe class imbalance in which rare pathologies are underrepresented and frequently misclassified, a limitation rarely exposed by the aggregate metrics reported in most prior work. We investigate a targeted intermediate-supervision framework, in which a secondary classifier head is attached to mid-level backbone features and jointly optimized with the primary classifier using inverse-frequency weighted loss. Unlike conventional deep supervision, which is primarily aimed at optimizing stability, the proposed formulation is used here to improve minority-class representation under severe OCT class imbalance. The method is evaluated on ResNet-18, ResNet-50, EfficientNet-B0, and ViT-B/16 using a four-class OCT dataset, with full per-class metrics reported across a systematic ablation of the auxiliary weight λ. EfficientNet-B0 achieved the best performance at λ = 0.3, attaining 97.78% accuracy, an AUROC of 0.995, and a Drusen F1-score of 93.51%, a gain of 2.64 percentage points over the unweighted baseline. Vision Transformers showed greater sensitivity to background padding artifacts than convolutional models. Grad-CAM and Attention Rollout analyses confirm that auxiliary supervision improves the localization of clinically relevant retinal structures, supporting its potential for interpretable, class-balanced automated OCT diagnosis.

Deep learning techniques have shown significant promise for the automated diagnosis of CVD using ECG analysis. Nevertheless, several critical challenges persist with current approaches: severe class imbalance, intricate temporal dependencies, and poor modeling of inter-beat relationships ultimately limit clinical applicability. This work presents a novel hybrid deep learning framework that combines a CNN for time-domain feature extraction, k-nearest neighbor-based hypergraph construction with cosine similarity to model inter-beat dependencies, and a residual-connection-enhanced hypergraph neural network (EHGNN) for robust classification. To address class imbalance, focal loss is implemented with adaptive class weighting. For evaluation, our model was tested on two benchmark datasets: the MIT-BIH Arrhythmia Database and the St. Petersburg Institute of Cardiological Technics (INCART) 12-lead Arrhythmia Database. In the case of the MITBIH dataset, a classification accuracy of 98.66% was achieved using the AAMI five-class classification system, whereas the results indicated a three-class arrhythmia detection accuracy of 95.46% for the INCART database. The model demonstrated consistent performance on two separate data sets, thus emphasizing its ability to generalize well. These results confirm that the proposed framework effectively addresses class imbalance through focal loss and SMOTE, temporal dependencies through CNN-based feature extraction, and inter-beat relationships through EHGNN, demonstrating great potential for clinical diagnostic systems by capturing complex heartbeat patterns that are otherwise overlooked by standard models.

BackgroundCervical cancer remains one of the leading causes of gynecological mortality worldwide, largely due to the limitations of manual cytological screening, which is time-consuming and susceptible to inter-observer variability. Although deep learning has demonstrated strong potential for automating cervical cytopathology, existing Convolutional Neural Network (CNN) models are hindered by two critical challenges: spatial irrelevance, where diagnostically meaningful nuclear regions are overshadowed by background artifacts such as blood and mucus, and severe class imbalance, where the dominance of normal cells impedes the accurate learning of rare dysplastic patterns.MethodsTo address these limitations, we propose DeepInsight-Net, a robust three-stage deep learning framework for cervical cell classification. The core architecture integrates Convolutional Block Attention Modules (CBAM) into a ResNet50 backbone to enhance spatial and channel-wise feature discrimination, enabling the network to selectively emphasize nuclear and nuclear–cytoplasmic boundary regions while suppressing irrelevant background noise. To further mitigate class imbalance, the conventional cross-entropy loss is replaced with Focal Loss, which dynamically down-weights easily classified samples and prioritizes hard, misclassified instances during training.ResultsExtensive experiments conducted on the benchmark SiPaKMeddataset demonstrate that DeepInsight-Net achieves a state-of-the-art classification accuracy of 99.63%, outperforming 15 competitive deep learning models, including EfficientNet-B6 and DenseNet169. Moreover, cross-dataset generalization experiments on an independent Liquid-Based Cytology (LBC) dataset yield an accuracy of 98.62%, confirming the robustness and domain adaptability of the proposed framework. Visual interpretability analyses using Grad-CAM and t-SNE reveal that the model consistently focuses on biologically relevant cellular regions, supporting the reliability of its predictions.ConclusionThe proposed DeepInsight-Net effectively addresses spatial irrelevance and class imbalance in cervical cytology analysis through attention-guided feature learning and loss re-weighting. The strong performance across multiple datasets, combined with transparent visual explainability, highlights its potential as a reliable computer-aided diagnosis (CAD) tool for supporting cervical cancer screening in real-world clinical settings.

Severe class imbalance between majority and minority classes in Big Data can bias the predictive performance of Machine Learning algorithms toward the majority (negative) class. Where the minority (positive) class holds greater value than the majority (negative) class and the occurrence of false negatives incurs a greater penalty than false positives, the bias may lead to adverse consequences. Our paper incorporates two case studies, each utilizing three learners, six sampling approaches, two performance metrics, and five sampled distribution ratios, to uniquely investigate the effect of severe class imbalance on Big Data analytics. The learners (Gradient-Boosted Trees, Logistic Regression, Random Forest) were implemented within the Apache Spark framework. The first case study is based on a Medicare fraud detection dataset. The second case study, unlike the first, includes training data from one source (SlowlorisBig Dataset) and test data from a separate source (POST dataset). Results from the Medicare case study are not conclusive regarding the best sampling approach using Area Under the Receiver Operating Characteristic Curve and Geometric Mean performance metrics. However, it should be noted that the Random Undersampling approach performs adequately in the first case study. For the SlowlorisBig case study, Random Undersampling convincingly outperforms the other five sampling approaches (Random Oversampling, Synthetic Minority Over-sampling TEchnique, SMOTE-borderline1 , SMOTE-borderline2 , ADAptive SYNthetic) when measuring performance with Area Under the Receiver Operating Characteristic Curve and Geometric Mean metrics. Based on its classification performance in both case studies, Random Undersampling is the best choice as it results in models with a significantly smaller number of samples, thus reducing computational burden and training time.

Automated disease screening systems face challenges when applied to multi-class medical image analysis, particularly under severe class imbalance inherent in clinical datasets. Retinal fundus imaging enables non-invasive screening for multiple ocular and systemic diseases simultaneously, yet current automated systems typically assess risk for only a single pathology or a limited disease range. We developed a comprehensive AI framework for simultaneous risk stratification of 15 distinct pathological conditions from retinal fundus images and report its preliminary evaluation in a real-world clinical setting. Our system combines fundus images from publicly available sources and proprietary clinical archives, addressing the significant class imbalance challenge inherent in rare condition risk assessment. We employed a hybrid deep learning architecture (CAFormer B36) with focal loss and targeted oversampling to ensure robust performance across common and rare conditions. The framework identifies retinal biomarkers associated with both primary ocular diseases (diabetic retinopathy, glaucoma, macular degeneration, cataracts, retinitis pigmentosa) and systemic conditions with retinal manifestations (hypertensive retinopathy, atherosclerotic changes, autoimmune manifestations). Dataset splitting was performed at the image level; all internal metrics should therefore be interpreted as exploratory upper-bound estimates pending patient-level replication. On the internal image-level test set, the system achieved ROC AUC of 0.9524-0.9971 across all 15 pathological classes, with F1 scores of 0.8968-0.9649. Notably, rare conditions with fewer than 100 training examples demonstrated robust risk stratification performance. In an exploratory single-site evaluation on 68 real-world cases, overall accuracy was 64.7% (95% CI: 52.9-76.5%). Due to the very limited number of sight-threatening cases in this small cohort (glaucoma n = 2, diabetic retinopathy n = 4), sensitivity estimates for these categories carry extremely wide confidence intervals and cannot be considered statistically reliable. These findings underscore the need for larger, prospective, multicenter studies. This pilot proof-of-concept demonstrates the feasibility of multi-pathology risk stratification under severe class imbalance using a hybrid deep learning architecture. Internal image-level metrics show strong discriminative capability, while the preliminary single-site evaluation (n = 68, 64.7% accuracy) reveals the challenges of real-world translation and the substantial gap that remains before clinical deployment. This work provides a methodological foundation for future prospectively validated screening systems.

ObjectiveThe imbalanced nature of real-world datasets is an ongoing challenge in the field of machine and deep learning. In medicine and in dentistry, most data samples represent patients not affected by pathologies, and on imagery, pathologic image areas are often smaller than healthy ones. Selecting suitable loss functions during deep learning is essential and may help to overcome the resulting imbalance. We assessed six different loss functions for one exemplary task, tooth structure segmentation on bitewing radiographs, for their performance. MethodsSix different loss functions (Focal Loss, Dice Loss, Tversky Loss and hybrid losses of Cross-Entropy and Dice Loss, Focal and Dice Loss, Focal and Generalized Dice Loss) were compared on a tooth structure segmentation task of 1,625 bitewing radiographs. Training was performed using three different model architectures (U-Net, Linknet, DeepLavbV3+) over a 5-fold cross-validation. Tooth structures consisted of the classes (occurrence in% of samples/captures areas measured on pixel level) enamel (100 %/25 %), dentin (100 %/50 %), root canal (100 %/10 %), filling (81 %/8 %) and crown (28 %/5 %). ResultsHybrid loss functions significantly outperformed standalone ones and provided robust results over the different architectures for the classes enamel, dentin, root canal and filling. Specifically, the Dice Focal loss reached high performance to conquer both image level and pixel level class imbalance, respectively. Clinical SignificanceIn dental use cases it is often important to predict minority classes such as pathologies accurately. Using specific loss function may be an effective strategy to overcome data imbalance when training deep learning models.

Cerebrovascular segmentation is important in various clinical applications, such as surgical planning and computer-aided diagnosis. In order to achieve high segmentation performance, three challenging problems should be taken into consideration: (1) large variations in vascular anatomies and voxel intensities; (2) severe class imbalance between foreground and background voxels; (3) image noise with different magnitudes. Limited accuracy was achieved without considering these challenges in deep learning-based methods for cerebrovascular segmentation. To overcome the limitations, we propose an end-to-end adversarial model called FiboNet-VANGAN. Specifically, our contributions can be summarized as follows: (1) to relieve the first problem mentioned above, a discriminator is proposed to regularize for voxel-wise distribution consistency between the segmentation results and the ground truth; (2) to mitigate the problem of class imbalance, we propose to use the addition of cross-entropy and Dice coefficient as the loss function of the generator. Focal loss is utilized as the loss function of the discriminator; (3) a new feature connection is proposed, based on which a generator called FiboNet is built. By incorporating Dice coefficient in the training of FiboNet, noise robustness can be improved by a large margin. We evaluate our method on a healthy magnetic resonance angiography (MRA) dataset to validate its effectiveness. A brain atrophy MRA dataset is also collected to test the performance of each method on abnormal cases. Results show that the three problems in cerebrovascular segmentation mentioned above can be alleviated and high segmentation accuracy can be achieved on both datasets using our method.

Imbalanced classification remains a critical challenge in decision-sensitive domains such as healthcare, finance, and cybersecurity, where minority class recognition is often paramount. This paper introduces FRSTU-Forest, a novel hybrid framework that integrates K-Nearest Neighbor (k-NN) imputation, Fixed Random State Undersampling (FRSTU), and Random Forest to enhance both minority class detection and model reproducibility. Unlike conventional undersampling, FRSTU applies deterministic sampling with a fixed random seed, ensuring consistent training subsets across runs and significantly reducing performance variance. The framework was comprehensively evaluated on seven benchmark datasets with moderate imbalance ratios (1.25–3.36) and rigorously tested on synthetic datasets with extreme imbalance ratios up to 1:100, high dimensionality (100 features), and substantial label noise (20%). FRSTU-Forest consistently outperformed baseline models (RF, k-NNimp+RF, RSTU+RF), achieving an average accuracy of 87.88%, minority-class F1-score up to 99.78%, and Cohen’s Kappa of 0.86 on benchmark datasets. More importantly, under extreme imbalance conditions (1:100 ratio), it maintained a balanced accuracy of 0.807 with 100 features, demonstrating remarkable robustness. Statistical significance was confirmed via the Bonferroni-Dunn test ( p = 0.0015 ), while computational efficiency remained practical with an average runtime of 2.01 seconds per dataset. In real-world applications, the model achieved 96.92% recall on breast cancer detection and substantially improved credit risk classification. These results affirm that FRSTU-Forest provides a reliable, reproducible, and robust decision-support tool for imbalanced data environments, particularly effective in scenarios with severe class imbalance, high dimensionality, and noisy labels.

• Multi-sine AC signals enable rapid, precise real-time PEMFC fault diagnosis. • SMOTE-Tomek balances and clears noisy data, sharply defining rare fault conditions. • Dual-attention WGAN synthesizes realistic PEMFC fault signals with high fidelity. • Achieves 99.67 % accuracy, surpassing state-of-the-art by 14 % at high imbalance. • Robust diagnosis and high accuracy using limited and imbalanced fault data. Proton exchange membrane fuel cell (PEMFC) faults, especially dehydration and flooding, cause distinct changes in electrochemical behavior. Consequently, real-time monitoring is essential for early and accurate diagnosis. However, acquiring real-world fault data is challenging, and the rarity of such faults results in severe class imbalance. This imbalance limits the performance, reliability, and practical applicability of conventional diagnostic methods. To address these limitations, this study proposes a unified diagnostic framework that integrates multi-sine AC voltage response, boundary-aware resampling, and attention-guided generative modeling. The key innovations of the proposed approach include: (1) Enhanced fault separability through the first application of multi-sine AC voltage response under data imbalance, enabling real-time extraction of critical electrochemical spectral features for early-stage diagnosis; (2) Improved data balance and clearer class boundaries using synthetic minority oversampling with Tomek links, which oversamples minority classes and removes borderline samples; (3) Realistic minority class synthesis using a dual attention Wasserstein generative adversarial networks, where channel attention focuses on diagnostically relevant spectral features and temporal attention models the dynamic evolution of PEMFC electrochemical behavior, ensuring high-quality, diagnostically informative synthetic fault data. The integrated framework achieves 99.67 % overall diagnostic accuracy and, under an extreme 1:200 class imbalance, outperforms state-of-the-art methods by 14 %. This approach enables rapid, data-efficient PEMFC fault diagnosis, strengthening fault management and advancing the performance of energy systems.

Gram-negative bacterial secreted effectors are translocated through specialized secretion systems to manipulate host cellular processes, and their accurate identification is crucial for understanding bacterial pathogenesis. Recent deep learning methods have significantly advanced this field, yet current approaches primarily rely on global sequence representations, overlooking the biological significance of terminal regions where secretion signals reside. Moreover, severe class imbalance among different secreted effector types remains a critical challenge for multi-class prediction. Here, we propose TermSE, a terminal signal-aware framework for multi-class secreted effector identification. TermSE explicitly captures N-terminal and C-terminal sequence features through convolutional neural networks applied to protein language model embeddings, and integrates them with global sequence representations for multi-view sequence characterization. To address class imbalance, TermSE employs a cosine-normalized classifier combined with weighted sampling to mitigate feature magnitude bias and ensure sufficient learning from minority classes. Extensive experiments demonstrate that TermSE outperforms existing methods in both cross-validation and independent test settings, with robust generalization across varying sequence identity levels. Furthermore, interpretability analysis confirms that TermSE learns to focus on biologically meaningful terminal patterns specific to each secreted effector type. These results highlight the potential of TermSE as an effective and interpretable tool for secreted effector discovery.

Although Deep learning (DL) methodologies have progressed substantially, the classification of skin cancer remains a challenging task. There are several reasons for this. Some of them are: artifacts like Hair; severe class imbalance in dermatoscopic datasets; and difficulty in extracting both fine-grained local features (details within small area of lesion) like texture, color, pigment network, vascular patterns and long range global features like overall shape, border irregularity, asymmetry. To overcome these, this study presents a novel, two-stage framework. At first, C’GAN (Conditional Generative Adversarial Network) is employed for generation of duplicate images for the minority classes. Then secondly, a CNN-ViT ensemble architecture is introduced followed by a cross attention based fusion module to fuse their features. The attention fusion model synergistically merges ViT’s global token representations with CNN’s local feature maps. The overall performance is analyzed through some standard quantitative metrics, whereas the reliability as well as the stability are validated through bootstrap based statistical analysis. The framework achieved remarkable accuracies of 99.3%, 99.7%, 98.9% and 98.2% on Dermatofibroma, Vascular lesions, Basal Cell Carcinoma, and Actinic Keratosis respectively, besides 99.4% overall AUC, 0.93 bootstrap mean, and 0.0003 standard error. The proposed model showed balanced performance both on majority as well as on minority classes, showcasing it’s effectiveness in class imbalance.

Stance detection has gained increasing interest from the research community due to its importance for fake news detection. The goal of stance detection is to categorize an overall position of a subject towards an object into one of the four classes: agree, disagree, discuss, and unrelated. One of the major problems faced by current machine learning models used for stance detection is caused by a severe class imbalance among these classes. Hence, most models fail to correctly classify instances that fall into minority classes. In this paper, we address this problem by proposing a hierarchical representation of these classes, which combines the agree, disagree, and discuss classes under a new related class. Further, we propose a two-layer neural network that learns from this hierarchical representation and controls the error propagation between the two layers using the Maximum Mean Discrepancy regularizer. Compared with conventional four-way classifiers, this model has two advantages: (1) the hierarchical architecture mitigates the class imbalance problem; (2) the regularization makes the model to better discern between the related and unrelated stances. An extensive experimentation demonstrates state-of-the-art accuracy performance of the proposed model for stance detection.

Drug-induced liver injury (DILI) presents a significant challenge due to its complexity, small datasets, and severe class imbalance. While unsupervised pretraining is a common approach to learn molecular representations for downstream tasks, it often lacks insights into how molecules interact with biological systems. We therefore introduce VitroBERT, a bidirectional encoder representations from transformers (BERT) model pretrained on large-scale in vitro assay profiles to generate biologically informed molecular embeddings. When leveraged to predict in vivo DILI endpoints, these embeddings delivered up to a 29% improvement in biochemistry-related tasks and a 16% gain in histopathology endpoints compared to unsupervised pretraining (MolBERT). However, no significant improvement was observed in clinical tasks. Furthermore, to address the critical issue of class imbalance, we evaluated multiple loss functions-including BCE, weighted BCE, Focal loss, and weighted Focal loss-and identified weighted Focal loss as the most effective. Our findings demonstrate the potential of integrating biological context into molecular models and highlight the importance of selecting appropriate loss functions in improving model performance of highly imbalanced DILI-related tasks.