Abstract
Background
Characterizing the tumor immune microenvironment (TIME) is essential for understanding anti-tumoral responses in colon cancer. This study introduces TIME Landscaper (TIMEL), a computational framework that uses deep learning to identify tissue structures at the microscopic level and summarizes their distribution across whole-slide images (WSIs) using statistical descriptors reflecting intratumoral heterogeneity for prognostic biomarker discovery.Methods
The performance of six deep learning image classification models (Inception V3, DenseNet-121, ViT-base, UNI, Prov-GigaPath, and Virchow) was evaluated to segment microarchitectural tumor, stromal, and immune areas using nearly 50,000 mini-patches. The selected model was applied to WSIs from discovery (TCGA-COAD; n = 411) and validation (Dartmouth; n = 108) cohorts, segmenting these components to map their spatial distribution. From these maps, 30 statistical descriptors representing abundance, variation, shape, spatial smoothness, and spatial heterogeneity were calculated as slide-level TIMEL features. Univariate and multivariate regression analyses identified survival-associated and metastasis-related biomarkers from the discovery and validation cohorts, respectively.Results
Virchow outperformed other models in segmenting tissue compartments (AUCs: 0.99, 0.98, 0.98 for tumor, stromal, immune components). Slide-level stromal and immune variance correlated with pathologist-assessed metrics (R = 0.47 and 0.32, both p p p = 0.001). Tumor spatial smoothness was associated with nodal metastasis (OR = 1.65; p = 0.02).Conclusion
TIMEL integrates deep learning and statistics to capture histological heterogeneity within the TIME, enhancing prognostic assessment and supporting precision oncology from routine histology.