Omni-Swin3D: A radiomics-aware architecture for robust lung cancer diagnosis

Benign–malignant classification of pulmonary nodules from chest computed tomography (CT) is essential for early lung cancer screening and for reducing unnecessary biopsies and follow-up examinations, yet it remains challenging because malignancy cues are subtle, heterogeneous, and distributed across 3D anatomy at multiple scales. We propose Omni-Swin3D, a radiomics-aware 3D hierarchical Transformer that preserves fine-grained texture and boundary evidence while enabling efficient global context modeling. Omni-Swin3D integrates a multi-scale radiomics encoder to capture complementary density/morphology patterns before tokenization, equips a four-stage Omni-Swin backbone with Adaptive Dual-Path Top-k Attention for sparse global aggregation with local continuity compensation, and adopts Adaptive Radiomics-Preserving Downsampling to mitigate discriminative information loss during hierarchical resolution reduction. A radiomics-aware prediction head further consolidates global average/max statistics via channel recalibration, and the model is optimized end-to-end with weighted cross-entropy to address class imbalance. Experiments on a public benchmark constructed from LUNA16/LIDC-IDRI demonstrate that Omni-Swin3D achieves 90.29±1.45% accuracy (balanced accuracy), 89.74±1.82% recall, 90.88±1.34% specificity, and an AUC of 0.980±0.008 under patient-level 10-fold cross-validation, while ablations confirm the complementary contributions and efficiency benefits of each component. Overall, these results indicate that explicitly preserving multi-scale radiomics evidence within an efficient 3D attention hierarchy yields robust malignancy discrimination. Future work will extend Omni-Swin3D toward multi-task prediction across multiple datasets by jointly modeling malignancy and additional clinically relevant targets.