SS 10 - DHI 1: Exploring the Future of AI, Radiomics, and Deep Learning in Radiation Oncology Research
159 - Explainable MRI-Based Deep Learning Model for Predicting EGFR Mutation in NSCLC Brain Metastases: Advancing Clinical Prognosis and Oncological Treatment Decisions
National Taiwan University Hospital Taipei, Taiwan
C. Y. J. Hsu1, W. Wang2, S. H. Kuo3, and H. H. Tsai.2; 1National Taiwan University Cancer Center, Taipei, Taiwan, 2National Taiwan University, Taipei, Taiwan, 3Department of Oncology, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei, Taiwan
Purpose/Objective(s): This study aims to develop and validate an explainable 2D/3D MRI-based deep learning model for predicting epidermal growth factor receptor (EGFR) mutation status in non-small cell lung cancer (NSCLC) brain metastases and to assess its oncological applications in predicting local tumor recurrence following stereotactic radiosurgery (SRS). Materials/
Methods: A retrospective analysis was conducted on a cohort of 463 NSCLC patients with brain metastases treated with radiosurgery. This cohort was divided into a primary group of 318 patients (759 brain metastases, treated between 2008 and 2020) and a validation group of 145 patients (345 brain metastases, treated between 2021 and 2022). The primary group was further split into a training set (252 patients, 616 brain metastases) and an internal validation set (66 patients, 143 brain metastases), maintaining a 4:1 ratio. We developed deep learning models using both 2D and 3D imaging inputs, processed through neural networks. The selection of the optimal network architecture (ResNet18, DenseNet121, SEResNet152, ViT) and hyperparameters was determined via five-fold cross-validation over 200 epochs. An explainable AI approach, utilizing occlusion sensitivity maps, was integrated to identify critical model features. The discriminative ability of the models was quantified using the area under the receiver operating characteristic curve (AUC). Competing risk analysis was employed to evaluate cumulative local recurrence risks. Results: In this study, SEResNet152 and ResNet18 were selected as the optimal models for 2D and 3D deep learning analyses, respectively. The ensemble 2D/3D deep learning model outperformed the individual models, achieving AUC values of 0.822, 0.760, and 0.755 for the training, internal validation, and temporal validation cohorts, respectively. Beyond traditional clinical predictors like age, gender, and smoking status, this 2D/3D ensemble model significantly correlated with EGFR mutation status, with odds ratios of 16.7 (95% CI: 10.8-26.5, p < 0.001) and 11.1 (95% CI: 5.42-24.3, p < 0.001) in the primary and temporal validation cohorts, respectively. By utilizing explainable AI through occlusion sensitivity maps, we identified Matched-OCC groups showing consistent map overlaps between the 2D and 3D models, which led to improved EGFR mutation prediction AUCs of 0.86, 0.82, and 0.82 across the cohorts. Furthermore, in predicting local tumor recurrence post-SRS, the ensemble model demonstrated superior prognostic ability compared to EGFR mutation status alone, with 1-, 2-, and 3-year AUC values of 0.605, 0.614, and 0.594, respectively, versus 0.577, 0.589, and 0.574 for EGFR mutation status. Conclusion: Our study demonstrates that 2D/3D MRI deep learning model and explainable AI effectively predicts EGFR mutation status and predicts local recurrence in NSCLC brain metastases treated with SRS. This approach offers a promising tool for personalized treatment planning and improving patient care.
