Cross-dataset adaptation of voxel-level deep radiomics for predicting survival in inoperable locally advanced NSCLC treated with immunotherapy

Jin Wang¹Zhaoyu Jiang^1,2Wenhao Ji¹Han Cheng^1,2Zhen Zhang^1,3,4Andre Dekker³Leonard Wee³Meng Yan^3,5*Xiaojing Lai^1,4*

• ¹Department of Radiation Oncology, Zhejiang Cancer Hospital；Hangzhou Institute of Medicine (HIM), Chinese Academy of Sciences; Zhejiang Key Laboratory of Particle Radiotherapy Equipment, Hangzhou, China
• ²Nanjing Medical University School of Public Health, Nanjing, China
• ³Department of Radiation Oncology (Maastro), GROW Research Institute for Oncology and Reproduction, Maastricht University Medical Centre+, Maastricht, Netherlands
• ⁴Department of Radiation Oncology, Key Laboratory of Cancer Prevention and Therapy, Tianjin Medical University Cancer Institute & Hospital, National Clinical Research Center for Cancer, Tianjin’s Clinical Research Center for Cance, Tianjin, China
• ⁵Department of Radiation Oncology, Key Laboratory of Cancer Prevention and Therapy, Tianjin Medical University Cancer Institute & Hospital, National Clinical Research Center for Cancer, Tianjin’s Clinical Research Center for Cancer, Tianjin, China

Background and purpose: Predicting overall survival (OS) for inoperable locally advanced non-small cell lung cancer (LA-NSCLC) treated with immune checkpoint inhibitors remains challenging due to heterogeneous clinical response. Furthermore, the application of advanced deep learning is hindered by limited immunotherapy datasets. This study aimed to develop a novel prognostic framework by integrating voxel-level deep radiomics derived from pretreatment imaging with a knowledge transfer strategy to accurately predict OS. Materials and methods: A total of 526 patients were respectively identified. A non-immunotherapy dataset from the RTOG 0617 clinical trial was used to pre-train a Vision-Mamba deep learning model to learn tumor characteristics within manually delineated tumor regions. Voxel-level radiomics feature maps were generated within tumors and integrated with CT images for dual-input co-training. Using the same dual-input, a cross-dataset transfer learning strategy was then used to adapt the pre-trained models to the immunotherapy context by fine-tuning. The model's performance was evaluated using the concordance index (C-index), time-dependent area under the receiver operating characteristic curve, Kaplan-Meier survival analysis, calibration curves, and decision curve analysis. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to suggest a possible interpretation of the model's decision logic. Results: The proposed model demonstrated robust generalization ability. In the independent immunotherapy testing dataset, the model achieved a C-index of 0.73 (95% CI:0.63-0.82). The time-dependent AUCs for predicting 1-year and 2-year OS were 0.73 and 0.70, respectively. Calibration curves showed good agreement between predicted and observed survival probability. Stratification analysis showed distinct survival differences, with the high-risk group exhibiting significantly poorer OS compared to low-risk group (P<0.001). Conclusion: We developed a voxel-level deep radiomics framework that bridges the data gap in immunotherapy research through fine-tuning on a limited immunotherapy dataset, and subsequent validation on an independent immunotherapy testing dataset, demonstrating robust generalizability.