Colorectal cancer (CRC) ranks among the most prevalent malignancies globally, with early detection and removal of precancerous polyps during colonoscopy being the most effective preventive strategy. However, conventional manual detection is hindered by operator fatigue and subjective assessment, resulting in clinically significant polyp miss rates of up to 27%. Existing computer-aided detection (CAD) systems face a fundamental trade-off: detection-only systems offer real-time speed but lack pixel-level precision, while segmentation-focused systems provide accurate delineation but are computationally prohibitive for live procedures. This paper presents DetSEG, a novel hybrid coarse-to-fine deep learning framework that resolves this trade-off through a two-stage architecture. In the first stage, a YOLOv8 object detection model performs rapid full-frame polyp localization, generating bounding box coordinates. In the second stage, cropped regions of interest are passed to a U-Net segmentation network with a pretrained ResNet-34 encoder for pixel-precise boundary delineation. We further propose an enhanced detection mode integrating Slicing Aided Hyper Inference (SAHI) with a dedicated small-polyp YOLOv8 model and an upgraded U-Net with EfficientNet-B4 encoder for improved detection of diminutive and flat polyps. The Kvasir-SEG dataset is utilized for model training and evaluation, supplemented by cross-dataset validation on CVC-ClinicDB. A comparative study against Faster R-CNN, SSD, and RT-DETR validates the superiority of YOLOv8 for the localization task. Experimental results demonstrate that DetSEG achieves a detection mAP_50 of 0.952, segmentation Dice coefficient of 0.891, and mean IoU of 0.842, with end-to-end inference under 100ms per frame. Additionally, a web-based CAD application with cloud deployment on Hugging Face Spaces is developed, providing an accessible clinical support tool. The proposed system substantially enhances colonoscopy procedures by offering both real-time detection and precise segmentation in a unified, deployable pipeline.
Fabric defect detection remains difficult when defects appear as subtle structural disruptions rather than strong color changes. This paper presents CAP-YOLO, a lightweight prior-coded input framework for YOLO-based fabric defect detection. Instead of redesigning the detector architecture, the method converts image-derived priors into RGB-compatible inputs so that multiple variants can be compared under the same YOLO26n detector, parameter count, training schedule, and data split. Two image-only priors are studied: a Delta E color-deviation prior and an anomaly-derived texture residual prior. Five variants are evaluated on the Tianchi fabric defect dataset: an RGB baseline (M0), Delta E prior coding (M1), anomaly prior coding (M2), dual-prior coding (M3), and a lightweight prior-adapter preprocessing route (M4). On the held-out test split, M2 achieves the best overall result with 0.3289 mAP50 and 0.1542 mAP50-95, improving over the RGB baseline by 18.48% in mAP50, 27.26% in precision, 9.38% in recall, and 16.62% in F1 score, while keeping the same detector size of 2.3787M parameters. Per-class analysis shows that the gains are concentrated in texture-disruptive defect categories, whereas categories dominated by diffuse appearance or subtle color spread benefit less. These results support a practical conclusion: for lightweight fabric defect detection, a simple image-derived texture residual prior is more effective than a color-only prior or naive multi-prior mixing under the current controlled setting.
Coronary Artery Disease (CAD) is one of the significant diseases in the field of cardiovascular science where timely identification of the risk will enable the diagnosis of CAD and its proper treatment. The advent of clinical data, laboratory data, and electrocardiographic data has led to the creation of artificial intelligence-based systems to predict CAD. Although traditional machine learning models are not very efficient in understanding the interactions among the features, on the other hand, the deep learning model provides good predictive capabilities but less interpretability. The existing CAD prediction approaches do not pay much attention to clinical transparency, feature interaction, calibration, and prediction explanation. The above-discussed problem makes the CAD prediction approaches unsuitable for the clinical environment. In this work, we propose a new Explainable Deep Cardiac Transformer (EDCT) that integrates deep feature learning, self-attention mechanism, residual feature refinement, and explainability for predicting CAD. As the proposed framework, we first preprocess and normalize the features followed by the deep cardiac feature embedding. The proposed transformer-based attention mechanism learns the interactions and long range dependencies between the heterogeneous clinical and cardiac features. Then, the obtained representations are fed into the classification layers to predict the CAD risk and identify the contribution of each clinical feature to the particular prediction using SHAP-based explanation analysis. According to the experiments, it was found that our proposed EDCT framework has outperformed all the existing approaches in all metrics. At the full training-data proportion, our EDCT framework gives 89.8% accuracy, 87.8% precision, 90.6% recall, 89.3% F1-score, and 0.938 ROC-AUC. Compared with XGBoost, the improvement of EDCT is 3.7%, 3.2%, 3.5%, 3.5%, and 0.032 for accuracy, precision, recall, F1-score, and ROC-AUC, respectively. Compared with AutoGluon Ensemble, the corresponding improvements are 2.5%, 2.1%, 2.4%, 2.4%, and 0.018, respectively. The experimental results show that the self-attention-based cardiac representation learning effectively captures interactions between the heterogeneous clinical and cardiac features.
Glaucoma is a progressive optic neuropathy that is among leading causes of irreversible blindness. Structural changes in the optic nerve head identification can assist in early clinical treatment. Retinal fundus photography is a non-invasive modality for glaucoma screening, while optic disc segmentation allows measuring morphological properties of clinical relevance. There are numerous deep learning algorithms developed for optic disc segmentation and glaucoma diagnosis; these tasks are often treated independently from each other. While convolutional architectures allow capturing the local spatial features effectively, they can experience challenges in capturing long-range dependencies and complex anatomy. Traditional segmentation networks can also suffer performance drop due to ambiguous, under-illuminated, occluded by blood vessels, or affected by pathology borders of the optic disc. This paper suggests Attention-Driven Vision Capsule Transformer (AVCT-Net) approach for simultaneous optic disc segmentation and glaucoma diagnosis. The approach involves hierarchical convolutional feature extraction, multi-scale attention mechanism, vision transformer-based global contextual analysis, and capsule-based feature representation. While the attention mechanism allows highlighting the most discriminating optic nerve areas, the transformer captures the long-range relationships between different retinal areas. Capsules provide preserving spatial and hierarchical dependencies between optic disc features and glaucoma-specific morphology. A multi-task learning approach is used for simultaneous optimization of segmentation and classification loss functions, thus, allowing utilizing anatomically relevant features generated by the segmentation branch in the diagnostic branch. According to experiments conducted on the REFUGE dataset, the proposed AVCT-Net algorithm allows achieving 98.5% accuracy, 97.7% precision, 98.2% recall, and 97.9% F1-score for glaucoma classification, as well as 98.4% Dice coefficient for optic disc segmentation within 100 training epochs. Compared with M-Net, Attention U-Net, and SwinCup-DiscNet approaches, the suggested approach allows improving accuracy by 2.7, 2.0, and 0.7%, respectively. The improvements in precision are 2.5, 1.8, and 0.6 points, while recall improvements are 2.1, 1.2, and 0.6 points, respectively. The improvements in F1-score are 2.3, 1.5, and 0.5 points, correspondingly. In optic disc segmentation task, AVCT-Net approach reaches 98.4% Dice coefficient, which is better than the results achieved by M-Net (by 2.1 points), Attention U-Net (by 1.4 points), and SwinCup-DiscNet (by 0.3 points).
Melanoma is an aggressive form of skin cancer in which early and accurate diagnosis substantially influences clinical management and patient survival. Dermoscopic imaging provides detailed visual information about lesion structures, pigmentation, borders, and internal patterns, making it an important source for computer-aided melanoma assessment. Recent deep learning methods have demonstrated strong diagnostic capability; however, their performance remains sensitive to image variability, class imbalance, subtle inter-class similarities, and limited representation of global lesion context. Conventional convolutional neural networks (CNNs) effectively capture local texture and morphological patterns but may inadequately model long-range relationships, whereas Transformer-based models capture global dependencies but can be sensitive to data availability and computational requirements. Existing hybrid models also frequently rely on simple feature concatenation without explicitly encouraging complementary and discriminative representations. This study proposes a novel Dual-Path Contrastive CNN-Transformer Network (DCCT-Net) for melanoma classification. The framework employs parallel CNN and Transformer pathways to learn complementary local and global lesion representations. A contrastive representation alignment module is introduced to reduce intra-class feature variation and increase inter-class separability, while an adaptive feature fusion mechanism combines the two pathways before classification. Augmentation-aware contrastive learning further improves representation robustness against changes in scale, illumination, rotation, and lesion appearance. The proposed DCCT-Net demonstrates superior performance compared with the DL-DNN, Vision Transformer, and Hybrid U-Net-Inception-ResNet-ViT approaches. At the final training epoch, DCCT-Net achieves 98.4% accuracy, 98.1% precision, 97.9% recall, 98.0% F1-score, and 99.0% specificity. Compared with the three existing methods, the proposed model improves accuracy by 2.2, 2.3, and 1.4%, respectively. The recall reaches 97.9%, representing improvements of 3.1, 3.4, and 1.9%, demonstrating stronger melanoma detection capability. The F1-score of 98.0% confirms a balanced improvement in precision and recall, while the specificity of 99.0% indicates effective discrimination of non-melanoma lesions. These results demonstrate that combining complementary CNN and Transformer representations with contrastive learning and adaptive fusion improves the discriminative capability of automated melanoma classification.
