Prognostic architecture of spatially resolved immune ecosystems in colorectal cancer: a multiplex immunohistochemistry and machine learning study

Purpose Colorectal cancer (CRC) prognosis is influenced by the tumor microenvironment (TME), including spatial immune cell organization and spatial immune aggregates (SIAs). However, the prognostic relevance of spatially resolved immune subsets, particularly B-cell-related phenotypes, remains incompletely defined and requires careful evaluation against established clinicopathological predictors. Methods A cohort of 786 CRC patients was analyzed using multiplex fluorescence immunohistochemistry (mIHC) targeting 10 immune phenotypes, coupled with deep learning-based single-cell segmentation and spatial statistics. A biologically informed density-based spatial clustering of applications with noise (DBSCAN) algorithm was used for automated SIA detection. Prognostic modeling was restricted to 636 patients with complete survival information and was evaluated using a leakage-free patient-level training/validation split. Benchmark models included TNM-based, clinicopathological, spatial-only, and integrated XGBoost models. Spatial heterogeneity, immune subset distributions, region-specific associations, and prognostic performance were systematically evaluated. Results Non-proliferating CD3 + CD8−FOXP3 − T cells dominated the TME (64.9%), and immune composition varied with disease stage and metastatic status. Automated SIA detection showed high concordance with pathologist-reviewed annotations and enabled region-specific profiling of immune phenotypes. SIA-region proliferating CD8 + T cells were associated with favorable outcome, whereas non-proliferating CD21 + B cells in non-SIA regions were associated with poorer outcome. Spatial proximity analysis identified metastasis-associated distance patterns, including shortened distances between CD21 − B cells and CD3 + CD8−FOXP3 − T cells. In prognostic benchmarking, clinicopathological variables remained the strongest predictors in the current cohort, whereas spatial features alone showed limited standalone discrimination. An integrated XGBoost model incorporating spatial and clinicopathological variables provided complementary prognostic information but did not materially outperform the clinicopathological benchmark. Conclusion Our study provides a high-resolution spatial immune atlas of CRC and identifies spatially resolved immune features associated with patient prognosis. The automated SIA detection framework and prognostic benchmarking analyses offer an exploratory basis for future validation of spatial biomarkers, while the current data support a complementary rather than dominant role for spatial features alongside established clinicopathological variables.