Cardiovascular diseases (CVD) remain the leading cause of global mortality, making early and accurate diagnosis essential for improving patient outcomes. However, most existing deep learning approaches address cardiac image segmentation or disease classification independently, limiting their effectiveness in complex clinical decision-making scenarios. In this study, we propose an explainable spatio-temporal deep learning framework that integrates segmentation-guided representation learning with efficient temporal modeling for automated CVD detection. The proposed architecture incorporates the Segment Anything Model for Medical Imaging in 2D (SAM-Med2D) to achieve accurate cardiac structure segmentation, followed by Mamba-based temporal feature extraction and Transformer-driven spatial representation learning to capture both dynamic motion patterns and anatomical dependencies in cardiac imaging sequences. To enhance transparency and clinical trust, Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) are employed to provide interpretable diagnostic insights. The framework is evaluated on three benchmark cardiovascular datasets, including EchoNet-Dynamic, CAMUS echocardiography, and UK Biobank cine cardiac magnetic resonance imaging (CMR). Experimental results demonstrate strong performance, achieving a Dice score of 91.20% for segmentation, an AUC of 95.50%, classification accuracy of 92.10%, and an MCC of 0.84, consistently outperforming multiple baseline methods. The proposed framework consistently outperforms baseline and existing methods, achieving approximately 3%–6% improvement in segmentation performance and 3%–4% improvement in classification accuracy across key evaluation metrics. The proposed approach offers a robust and explainable solution for automated cardiovascular disease detection, with significant potential to support reliable clinical deployment and improve diagnostic workflows in medical imaging practice.

Hyperspectral image (HSI) denoising is a crucial preprocessing step that significantly enhances the performance of downstream applications, such as object detection and classification. Whereas deep neural networks have achieved remarkable performance in HSI denoising, many existing models rely mostly on vanilla convolutions, which often fail to capture fine-grained noise patterns and structural details in real-time HSIs. To address these limitations, we propose a novel Center-Difference Convolutional Network (CDCN) designed to effectively suppress various noise types while preserving the inherent structure of HSIs. By leveraging center-difference convolution (CDC), our model captures both gradient and intensity information in the spatial domain, enabling better discrimination of subtle noise characteristics. The CDCN architecture processes 3D HSI cubes through separable 3D convolutions, efficiently extracting spatial-spectral features with minimal computational overhead. Additionally, a spatial-spectral attention mechanism is integrated to further refine feature representation. We evaluate the proposed method on one simulated dataset (Kennedy Space Center) and two real-world datasets (Pavia Center and Houston-2018). Experimental results demonstrate that CDCN consistently outperforms existing state-of-the-art approaches, achieving superior denoising performance while maintaining spectral-spatial information. Ablation studies also validate the effectiveness of CDC and attention mechanisms in enhancing denoising capability over standard convolutional baselines.