DSR-VGG: A Lightweight Depthwise Separable Residual VGG Architecture for Nasopharyngeal Histopathological Image Classification

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.