Interpretable deep learning and radiomics for predicting the occurrence of brain metastases following prophylactic brain irradiation in small cell lung cancer.
Authors
Affiliations (5)
Affiliations (5)
- School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai, China.
- Department of Radiation Oncology, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- Department of Logistics, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- Department of Science and Education, Shanghai Jing'an District Mental Health Center, Shanghai, China.
- Department of Shanghai Lung Cancer Center, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
Abstract
Brain metastasis remains an important clinical concern in patients with small cell lung cancer (SCLC), even after prophylactic cranial irradiation (PCI). Accurate identification of patients at high risk of brain metastasis after PCI may facilitate individualized surveillance and management. This study developed a fusion model based on two-and-a-half-dimensional (2.5D) multi-channel deep learning features, clinical features, and radiomics features to predict whether patients with SCLC will develop brain metastases following PCI. This retrospective study included 185 patients with SCLC who underwent PCI and randomly divided them into a training set (N=148) and a testing set (N=37). A 2.5D multi-channel convolutional neural network (CNN) model was constructed based on multi-planar regions of interest (ROI) derived from CT images to extract 2.5D deep learning features. Based on the largest cross-section and three-dimensional (3D) volume of the tumor ROI, two-dimensional (2D) and 3D deep learning features were extracted, respectively. Subsequently, features from different dimensions were fused with clinical features and traditional omics features to construct the 2D_clinic_Rad, 2.5D_clinic_Rad, and 3D_clinic_Rad models, respectively. The performance differences between the 3D models and the clinical and traditional radiomics models were then compared. The 2.5D_clinic_Rad fusion model demonstrated the best performance in predicting brain metastases, achieving an area under the curve (AUC) of 0.841 on the test set, outperforming clinical models, traditional radiomics models, and 2D and 3D deep learning fusion models. Analysis of decision curves and calibration curves indicated that the 2.5D_clinic_Rad model demonstrated the highest clinical net benefit and the best calibration consistency compared to other models. The fusion model based on 2.5D_clinic_Rad can effectively predict the risk of brain metastasis in patients with SCLC following whole-brain prophylactic irradiation.