Quantitative radiomics-based biomarker discovery for prognostic stratification in gastroesophageal cancer: Insights from the Cancer Moonshot Biobank Imaging Cohort.

Abstract

438 Background: Radiomics has emerged as a promising approach for non-invasive biomarker discovery in oncology. However, its application in gastroesophageal cancer (GEC) remains underexplored. Leveraging the Cancer Moonshot Biobank Imaging Cohort (CMB-GEC), we evaluated whether longitudinal CT imaging could yield reproducible radiomic biomarkers with potential prognostic relevance. Methods: We analyzed CT series from 6 patients with GEC (CMB-GEC). DICOM studies were converted to isotropic 1-mm NIfTI volumes using SimpleITK, followed by semi-automated tumor segmentation. Quantitative radiomic features (~1,500 per case; intensity, texture, and shape) were extracted using PyRadiomics. Features were cleaned by excluding diagnostic metadata, constant values, and all-NaN columns, yielding a stable feature set. Heatmaps and correlation analyses were performed to identify feature redundancy and patient clustering. Results: Initial extraction: ~1,500 features per patient. Stable features retained after cleaning: 6 shape/size descriptors (e.g., VoxelVolume, Major/Minor axis lengths, Elongation, Flatness). Strong correlations were observed among volume and axis features (ρ > 0.9). Elongation and Flatness were moderately correlated (ρ ≈ 0.69). Patient clustering revealed two dominant phenotypes: Size-driven tumors (MSB-04437, MSB-01493); shape-driven, elongated tumors (MSB-06857, MSB-09348); two patients showed intermediate profiles. Conclusions: This pilot radiomics analysis demonstrates that size and shape-related biomarkers can be stably extracted from CT imaging in GEC, even in a small cohort. While most advanced texture features collapsed due to limited sample size, clustering revealed biologically plausible subgroups. Expansion to the full CMB-GEC cohort will enable validation and exploration of prognostic associations with treatment outcomes.