Background: Accurate differentiation of kidney diseases such as cysts, tumors, stones, and normal tissue from computed tomography (CT) images remains challenging due to overlapping visual characteristics and variability in data distributions. While deep learning approaches have shown promising results, many existing studies rely on image-level data splitting and focus primarily on accuracy, which may lead to overly optimistic performance and limited clinical reliability. Methods: This study proposes KDH-Net (Kidney Disease Hybrid Network), a hybrid deep learning framework for multiclass kidney disease characterization that integrates EfficientNetB0, ResNet50, and MobileNetV2 through feature-level fusion. A two-stage training strategy is adopted to enhance optimization stability. To ensure realistic performance assessment, experiments on the primary dataset are conducted under a patient-level evaluation protocol, eliminating potential data leakage. The framework further incorporates calibration analysis, statistical validation, and explainable artificial intelligence to evaluate prediction reliability and interpretability. Results: On the patient-level dataset, KDH-Net achieves an overall accuracy of 0.93 with a macro-average F1-score of 0.91, demonstrating balanced performance across all classes. Confidence analysis indicates meaningful alignment between prediction confidence and correctness, while Grad-CAM visualizations highlight anatomically relevant regions associated with each class. Conclusions: The results demonstrate that KDH-Net provides a stable, reliable, and interpretable framework for kidney CT characterization. The proposed system is designed to support clinical decision-making by offering trustworthy predictions under realistic evaluation conditions, rather than replacing clinical expertise.

Background: Nasopharyngeal carcinoma (NPC) presents significant diagnostic and therapeutic challenges, often due to late-stage detection and its complex anatomical location. The increasing integration of artificial intelligence (AI) into oncology offers potential opportunities to enhance the precision of NPC management. This systematic review aims to synthesise the current evidence of AI applications in NPC diagnosis, prognostication, and treatment planning. Methods: A systematic literature search was conducted following PRISMA guidelines across multiple databases (PubMed, Scopus, Embase, Google Scholar, IEEE Xplore) for studies published up to June 2025. From an initial pool of 2549 articles, 55 studies meeting the inclusion criteria were selected for qualitative analysis. The review focuses on AI models applied to key diagnostic modalities: computed tomography (CT), magnetic resonance imaging (MRI), and histopathological whole-slide images (WSI). Results: AI, particularly deep learning (DL), shows promising performance in automating critical tasks across all modalities. For CT and MRI, models have been reported to achieve accurate tumor and organ-at-risk segmentation, potentially supporting radiotherapy planning, and show strong performance in predicting survival outcomes and treatment toxicity. In digital pathology, AI enables automated diagnosis and facilitates the extraction of prognostic “pathomic” features from WSIs, with some studies suggesting performance comparable to or exceeding traditional radiomics. The most significant advances are seen in multimodal AI systems that integrate radiological, pathological, and clinical data, which, in some studies, show modest improvements in prognostic performance compared to single-modality approaches. However, these findings are preliminary, as none of the reviewed multimodal models underwent rigorous external validation in large, multi-center cohorts. Reported performance varies considerably across studies, and claims of superiority should be interpreted with caution.

Background/Objectives: Deep convolutional neural networks achieve strong histopathological image-classification performance but often require substantial computational resources. This study proposes DSR-VGG, a lightweight Depthwise Separable Residual VGG architecture for nasopharyngeal histopathological image classification. Methods: DSR-VGG comprises six Depthwise Separable Blocks with residual projection shortcuts, global average pooling, dropout regularization, and a classification layer. An inverse-frequency weighted random sampling strategy was used to address severe class imbalance. The primary evaluation used 88,002 nasopharyngeal histopathology patches from four classes: Normal, lymphoid hyperplasia, nasopharyngeal inflammation, and nasopharyngeal carcinoma. Training was repeated using three random seeds, with supplementary public datasets used to assess cross-dataset architectural robustness. Results: Across three runs, DSR-VGG achieved 99.764 ± 0.049% accuracy, 99.070 ± 0.365% balanced accuracy, and 99.219 ± 0.215% macro-F1. The model contained only 0.528 million parameters, required approximately 0.31 GFLOPs, and achieved 2.38 ms single-image inference latency under the stated GPU benchmarking conditions. Although several baseline architectures achieved marginally higher accuracy, DSR-VGG maintained competitive predictive performance with substantially lower computational complexity. Conclusions: DSR-VGG provides a favorable accuracy–complexity trade-off for histopathological patch classification. However, patient-/WSI-disjoint evaluation, independent multi-center validation, and whole-slide assessment remain necessary before clinical deployment.

Deep convolutional neural networks have achieved remarkable success in histopathological image classification; however, many state-of-the-art architectures rely on deep and computationally expensive designs that limit deployment in resource-constrained clinical environments. Inspired by the hierarchical feature extraction strategy of VGG networks, this study proposes DSR-VGG, a lightweight Depthwise Separable Residual VGG architecture for nasopharyngeal histopathological image classification. Unlike conventional CNN architectures that repeatedly stack multiple standard convolutional layers within each stage, the proposed network employs a single Depthwise Separable Block (DSBlock) per stage, where standard convolutions are replaced by depthwise-pointwise separable convolutions to substantially reduce computational complexity. To preserve feature propagation and improve optimization stability, residual projection shortcuts are integrated throughout the network. The proposed architecture consists of six DSBlocks followed by global average pooling, dropout regularization, and a softmax classifier. A weighted random sampling strategy is further employed to mitigate severe class imbalance in the training dataset. The primary clinical evaluation was conducted on a large-scale multi-center nasopharyngeal biopsy dataset comprising 88,002 image patches from four diagnostic categories. To assess the architectural robustness of DSR-VGG beyond its primary NPC application, supplementary experiments were also conducted on several publicly available medical image-classification benchmarks. These additional experiments were used to evaluate cross-dataset robustness and computational efficiency; they were not intended as external clinical validation of the NPC classification task. The proposed model achieved 99.8% classification accuracy on the NPC dataset while utilizing only 0.534 million parameters, 0.31 GFLOPs, and an inference latency of 2.38 ms. Ablation and comparative analyses demonstrate that the proposed architecture achieves competitive or superior performance relative to conventional and lightweight deep learning models while maintaining a significantly smaller computational footprint. These findings highlight the potential of DSR-VGG as an efficient and deployable solution for real-time computer-aided diagnosis in digital pathology.