Explainable Deep Survival Learning Reveals Tumor Spatial Features and Stromal Patterns for Recurrence-Free Survival Prediction In Neuroendocrine Tumor Liver Metastases
Abstract
Purpose
To develop an explainable deep learning framework for histological segmentation and prognostic modeling of neuroendocrine tumors (NET) liver metastasis, comparing the efficacy of non-linear deep survival analysis against traditional linear Cox regression.
Methods
A Vision Transformer (ViT-L/16) model was trained on 120 WSIs (4:1 split) and integrated with a sliding window algorithm for semantic segmentation. We algorithmically defined satellite tumors—micro-tumor nests separated from the primary mass by normal hepatic parenchyma—using connected component analysis. High-dimensional spatial features, including the 95% Hausdorff Distance (HD95), Specific Surface Area (SSA), and cancer-associated fibrosis (CAF) ratios, were extracted. Prognostic modeling was conducted on 131 patients (513 WSIs) partitioned into training (n=91) and validation (n=40) cohorts. A univariate Cox analysis screened candidate features (p<0.1), which were subsequently evaluated via a multivariate Cox model. We compared a traditional multivariate Cox model with an explainable DeepSurv network utilizing layer-wise relevance propagation (LRP) to evaluate non-linear fitting capabilities and feature robustness.
Results
The ViT-L/16 achieved a validation accuracy of 91.77%.In univariate analysis, five features emerged as significant (p<0.1), led by satellite HD95 mean (HR=1.58, p=0.006) and sattellite count (HR=1.44, p=0.01). In survival analysis, the multivariate Cox model identified satellite HD95 as a significant predictor (HR=1.90, p=0.0026) but yielded moderate discriminative power (Val C-index: 0.6069). Conversely, the DeepSurv model demonstrated superior non-linear fitting and noise immunity, significantly improving performance (Val C-index: 0.7645). LRP interpretability confirmed that main tumor area, satellite HD95, and CAF-fibrosis-to-blood ratios were the most stable morphological drivers of risk.
Conclusion
Deep survival modeling with LRP provides a more robust and efficient prognostic tool for NETLM than linear frameworks. By capturing complex non-linear interactions, this explainable AI approach identifies tumor-stroma spatial features and stromal patterns as stable digital biomarkers for precise risk stratification of tumor recurrence.