2179 - Uncovering Tumor Heterogeneity Captured by MRI-Based Radiomic Features across Clinical and Outcome Subgroups in Pancreatic Ductal Adenocarcinoma

Purpose/Objective(s):

Pancreatic ductal adenocarcinoma (PDAC) is an aggressive malignancy characterized by substantial inter- and intra-tumoral heterogeneity, and inconsistent clinical outcomes following MRI-guided stereotactic body radiotherapy (MRgSBRT). MRI-based radiomic features may capture tumor phenotype and imaging heterogeneity associated with treatment response. This study evaluated whether radiomic features differ across PDAC subgroups defined by tumor location, grade, failure patterns, and survival to inform future predictive model development.

Materials/Methods:

We conducted a retrospective study of 106 patients with PDAC treated with SBRT on a 0.35 T MRI linac between 2015 and 2024. Comprehensive clinical, treatment, and outcome data were collected. Quantitative imaging features were extracted from the gross tumor volume (GTV) on planning 0.35 T TrueFISP MRI using the Computational Environment for Radiotherapy Research (CERR). A total of 332 features, including shape, first-order intensity, and texture-based features, were analyzed. Imaging heterogeneity was evaluated using the coefficient of variation (CV) and relative CV difference (?CV) across binary subgroups defined by tumor location (head vs body/neck/tail), grade (well/moderately differentiated vs poorly differentiated), and clinical endpoints, including failure patterns (in-field vs out-of-field), and survival interval (< 1 year vs ³ 1 year from the start of RT). Statistical comparisons were performed using the Wilcoxon rank-sum and Welch two sample t-tests, with a significant threshold of a <= 0.05).

Results:

Radiomic heterogeneity and statistical differences varied across subgroups. The greatest number of significantly heterogeneous features was observed between patients with less than one year overall survival after SBRT and those who lived greater than one year (60 by rank-sum; 41 by t-test), with 27 features demonstrating ?CV > 0.30. First-order kurtosis showed the strongest intergroup difference (?CV = 0.77, p = 0.001), followed by first-order skewness and GLCM maximum inverse difference normalized. In-field versus out-of-field failure were found to be different in 27 significant features (p<0.05), whereas in-field failure alone showed limited differences. A difference in tumor location (head/neck/body/tail) demonstrated 20 (rank-sum) and 10 (t-test) significant features, with first-order kurtosis (?CV =0.58, p = 0.004), and shape flatness (?CV = 0.40, p = 0.004) among the most distinct. Tumor grade exhibited substantial heterogeneity differences (187 features with ?CV > 0.30) despite fewer statistically significant features.

Conclusion:

Radiomic features demonstrated statistically significant differences across clinically relevant PDAC subgroups. These findings suggest that low field MRI-based radiomics capture meaningful tumor phenotypic variation and may inform future outcome prediction and risk stratification to support personalized and adaptive strategies.