Exploratory associations between regional immune-stromal features and 18F-FDG PET-derived metabolic heterogeneity in breast cancer

Abstract Background The breast tumor microenvironment harbors spatially heterogeneous immune and stromal populations that may influence functional imaging. This exploratory study examined whether regional immune-stromal differences relate to 18 F-FDG PET-derived metabolic heterogeneity and radiomics phenotypes, and which metrics show the largest nominal correlations. Methods Retrospective study of 61 women with invasive breast carcinoma undergoing baseline 18F-FDG PET/CT. A pathologist blinded to PET annotated ten intratumoral regions (≥ 1 mm²) per case on whole-slide images; QuPath machine-learning classification quantified stromal tumor-infiltrating lymphocytes (sTILs) and fibroblast-like stromal cells (FLSC). The regions with numerically highest and lowest sTIL density were retained for paired analysis; values therefore represent the observed range, not a comprehensive spatial map. The H&E-based FLSC measurement was cross-checked against α-smooth muscle actin (α-SMA) immunohistochemistry in the same regions as a supporting measurement check. PET features comprised SUVmax, MTV, TLG, texture (GLCM, GLRLM, GLSZM) and heterogeneity indices (HI2, HI3; MTV30–MTV50 slope). Spearman correlation with Benjamini–Hochberg FDR correction was used. Results Of 270 correlations (nine immune-stromal endpoints × 30 PET features), 38 reached nominal significance ( p < 0.05), but none survived FDR correction within an endpoint (smallest q = 0.077) or study-wide (smallest q = 0.33). The largest coefficient was weak: |ρ|=0.379 (TIL-high region sTIL density vs. HI2; 95% CI 0.126–0.586; p = 0.0026). sTIL density correlated inversely with MTV (ρ=−0.29 to − 0.35) and positively with HI2; the TIL/FLSC ratio correlated with SUV parameters (largest ρ = 0.299, SUVmean). FLSC density and regional difference metrics were weaker. Across 122 paired regions, α-SMA-based stromal density agreed closely with the H&E-based measurement (ρ = 0.992; Lin CCC = 0.996), and reanalysis reproduced the same correlation profile ( r = 0.94–0.996). Excluding extreme densities or smallest regions, 9–24 of 38 associations persisted; after adjustment for T stage, Nottingham score, Ki-67 and subtype, 13 persisted. Conclusions This hypothesis-generating study presents a paired, region-level integration of machine-learning digital pathology with PET radiomics in breast cancer. Only weak nominal associations between regional immune-stromal composition and PET-derived metabolic heterogeneity were observed, and none remained significant after FDR correction. These findings require independent validation in larger, multi-institutional cohorts before any biological or clinical interpretation can be drawn.