| 摘 要: 针对非小细胞肺癌腺癌/鳞癌亚型判别中三维CT标注样本不足及影像—临床融合不充分问题,提出可解释多模态框架CGS-Mamba。基于240例未标注CT以3D-MAE进行补丁重建预训练获取三维特征提取器,并将CEA、年龄、性别与TNM分期映射为Mamba选择性参数以实现临床条件化特征提取,结合Cross-modalTransformer完成影像Patch与临床Token对齐交互。在103例标注数据上,内部测试AUC为0.932,外部验证AUC为0.898,优于3D-ResNet18、Swin-ViT及常见融合策略。CEA重要性最高,表明方法具备临床可解释性并提升亚型分类精度。 |
| 关键词: 非小细胞肺癌 3D-MAE Mamba 临床引导 多模态融合 |
|
中图分类号: TP391.4
文献标识码: A
|
|
| Interpretable Non-Small CellLung Cancer Subtype Classification Based on 3D-MAEand Clinical-guided Selective Scanning(CGS-Mamba) |
|
WANG Xiaotong, QIAN Qian, XIA Tian, HAN Lei, SUN Liping
|
(School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China)
wxt857455412@ 163.com; qianq9720@ 163.com; 243352316@ st.usst.edu.cn; 1264883510@ qq.com; sunlp@ Sumhs.edu.cn
|
| Abstract: To address the scarcity of labeled 3D CT data and insufficient imaging-clinical fusion in non-small cell lung cancer (NSCLC) adenocarcinoma/squamous subtype classification, an interpretable multimodal framework, CGS- Mamba, is proposed. A 3D feature extractor is obtained via 3D-MAE patch-reconstruction pretraining on 240 unlabeled CT scans. Clinical variables (CEA, age, sex, and TNM stage) are mapped to selective parameters of a Mamba state- space model to enable clinically conditioned feature extraction, and a cross-modal Transformer is employed for fine- grained interaction and alignment between imaging patches and clinical tokens. Experiments on 103 labeled cases achieve an AUC of 0.932 on the internal test set and 0.898 on an external validation set, outperforming 3D-ResNet18, Swin-ViT, and common fusion strategies. Feature-importance analysis indicates that CEA contributes most to the prediction, supporting clinical interpretability while improving subtype classification performance. |
| Keywords: NSCLC 3D Masked AutoEncoder Mamba clinical guidance multimodal fusion |
