SCLC-SBT: a spectral-bilinear transformer for small cell lung cancer histopathology image classification

Small cell lung cancer (SCLC) is among the most lethal thoracic malignancies, and timely and accurate histopathological classification is critical for guiding treatment decisions. Existing deep learning approaches face difficulty simultaneously capturing the high-frequency nuclear texture features characteristic of SCLC—such as hyperchromatic salt-and-pepper chromatin and dense nuclear moulding—and the long-range spatial dependencies of tissue organisation, limiting classification reliability. To address these challenges, this study proposes SCLC-SBT, a hierarchical Spectral-Bilinear Transformer for automated SCLC histopathological image classification. The network deploys re-parameterizable Hybrid-Conv Blocks (HCB) with frequency-domain Spectral-Channel Attention (SCA) in the first three stages to extract fine-grained multi-scale local features, and employs a novel Dual-Branch Conv Linear Integration Transformer (DBCLIT) in the deepest stage to jointly model local morphology and long-range dependencies via bilinear selective state space modelling. Evaluated on a 1,974-image binary classification benchmark comprising Normal and SCLC tissue patches, SCLC-SBT achieves 0.9697 accuracy (95% CI 0.9495 to 0.9865), 0.9728 F1-score, and 0.9971 AUC, attaining the best point estimate among eleven compared architectures including DeiT, MLP-Mixer, DaViT, and CAFormer on five of the seven reported metrics. Bootstrap confidence intervals and pairwise McNemar’s tests indicate that this advantage is statistically significant relative to five of the eleven baselines. Relative to the six strongest baselines, DeiT, MLP-Mixer, CAFormer, Swin-T, and the tied PoolFormer and DaViT pair, the difference is not statistically significant at α =0.05 on the 297-sample test set, so SCLC-SBT is best interpreted as achieving competitive to superior accuracy at a substantially better accuracy-efficiency trade-off while maintaining competitive model efficiency with approximately 17 M parameters and 7.0 G FLOPs. The results demonstrate that integrating spectral-domain channel recalibration with bilinear state space modelling provides an effective inductive bias for SCLC histopathological feature learning.