3680 - Whole Slide Image-Based Pathomics Fully Supervised Deep Learning Model and Postoperative Recurrence-Free Survival in Stage IA Non-Small Cell Lung Cancer
Presenter(s)
X. Yin1, Y. Lu2, Y. Cui1, R. Yan2, Y. Gao3, Z. Huang4, J. Yu5, and X. Meng1; 1Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China, 2School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China, 3Department of Pathology, Shandong Cancer Hospital Affiliated to Shandong University, Shandong Academic of Medical Science, Jinan, China, 4Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan, China, 5Shandong Provincial Key Laboratory of Precision Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, Shandong, China
Purpose/Objective(s):
Up to 30% of patients experience tumor recurrence after standard surgery. While clinical guidelines recommend adjuvant therapy for high-risk patients with stage IB–IIIA NSCLC, evidence supporting routine adjuvant therapy remains insufficient for high-risk stage IA patients. This unmet clinical need highlights the urgency of developing accurate recurrence prediction tools for stage IA NSCLC. This study aims to construct a pathomics model utilizing whole-slide imaging (WSI) to predict recurrence in patients with surgically resected stage IA NSCLC.Materials/Methods:
This retrospective study included 293 patients with stage IA NSCLC undergoing radical resection. We constructed a spatially structured pathological prediction framework integrating pixel-level tissue segmentation, regional phenotyping, and graph neural network (GNN) for recurrence prediction. A U-Net++ model was trained using 100 pathologist-annotated WSIs covering 9 tissue types to achieve automatic segmentation. Image patches were organized as spatial graphs, and GNN was employed to capture spatial dependencies within the tumor microenvironment. Finally, recurrence risk prediction and survival stratification were realized via graph embedding and fully connected layers.Results:
A total of 293 patients were randomly assigned to the training and validation cohorts at a ratio of 8:2. The median follow-up time was 63.86 months, and 16.4% of patients (n = 48) experienced recurrence. Distant metastasis was the predominant first recurrence pattern (52.08%), followed by locoregional recurrence (25.00%) and combined locoregional plus distant recurrence (22.92%). Using 5-fold random cross-validation, the model achieved the area under the curve (AUC) of 0.976 and accuracy (ACC) of 0.947 in the training set, and an average AUC of 0.800 and average ACC of 0.763 in the validation set.Conclusion:
By explicitly modeling spatial topology for structured relational learning, our study outperforms conventional multiple instance learning (MIL)-based feature aggregation, yielding improved prognostic performance and model interpretability. It provides a precise prognostic tool for individualized postoperative therapy in stage IA NSCLC and may further improve outcomes in early-stage lung cancer through personalized follow-up and treatment stratification.