PCMamba: Integrating partial convolution and state space models for lung cancer histopathological image classification

Accurate classification of lung cancer histopathological images is essential for diagnosis and treatment planning, yet existing approaches face challenges in effectively capturing both local texture patterns and long-range spatial dependencies. Convolutional neural networks excel at local feature extraction but exhibit limited long-range dependency modeling. Transformer-based methods capture global dependencies effectively but require substantial computational resources. This paper proposes PCMamba, a novel architecture integrating Partial Convolution (PConv) with Selective State Space Models (SSMs) for lung cancer histopathological image classification. PCMamba employs a dual-branch design where PConv captures local texture patterns by convolving only a subset of channels, while SSMs model long-range spatial dependencies with linear complexity. Extensive experiments on the LungHist700 dataset demonstrate superior performance: 94.17% accuracy, 94.42% F1-score, and 99.28% AUC, outperforming 11 representative baselines under our experimental settings. Ablation studies validate each component's contribution, sensitivity analyses confirm robustness across hyperparameter variations, and Grad-CAM visualizations demonstrate clinically relevant feature focus. PCMamba provides an effective approach for automated lung cancer diagnosis from histopathological images.