3502 - Interpretable 2.5 D Multi-Channel Deep Learning for Predicting Brain Metastasis after Prophylactic Cranial Irradiation in Small Cell Lung Cancer
Presenter(s)
Y. Huang; Shanghai Chest Hospital, Shanghai, Shanghai, China
Purpose/Objective(s):
This study developed a fusion model based on 2.5D multi-channel deep learning, clinical, and radiomics features to predict brain metastasis in small cell lung cancer (SCLC) patients following prophylactic cranial irradiation (PCI).Materials/Methods: A retrospective analysis was conducted on 185 SCLC patients who received PCI, randomly divided into training (N=148) and test (N=37) sets. A 2.5D multi-channel convolutional neural network (CNN) model was constructed using multi-view ROI from CT images to extract 2.5D deep learning features. These features were fused with clinical and traditional radiomics features to build the 2.5d_clinic_Rad model. Machine learning algorithms were employed for classification prediction, and performance was compared with clinical, radiomics, 2D, and 3D deep learning fusion models. Additionally, Grad-CAM was used for visualization and SHAP for feature contribution analysis to ensure model interpretability.
Results:
The 2.5d_clinic_Rad fusion model demonstrated superior performance in predicting brain metastasis, achieving an AUC of 0.851 (95% CI: 0.7080-0.9944) in the test set, significantly outperforming clinical (AUC 0.605), radiomics (AUC 0.599), 2D (AUC 0.819), and 3D (AUC 0.710) models. Grad-CAM visualization effectively localized key lesion areas in CT images, while SHAP analysis revealed that 2.5D deep learning features contributed more significantly to prediction than traditional radiomics features.Conclusion:
The 2.5d_clinic_Rad multi-channel fusion model effectively predicts brain metastasis risk in SCLC patients after PCI. The interpretability analysis enhances the clinical credibility of AI-assisted decision-making.