ESTRO 2025 - Abstract Book

S449

Clinical - Breast

ESTRO 2025

was applied to identify representative features. In the prognostic analysis cohort, which contained longitudinal ADC images and clinical records, the T0 representative feautures and their percentage changes at early-treatment (ΔT1) and mid-treatment (ΔT2) were analyzed using univariate and multivariate methods to evaluate correlations with pCR. Established prognostic factors, including hormone receptor (HR), humen epidermal receptor 2 (HER2), and percentage change in mean ADC at mid-treatment (ΔT2 ADCmean), were used as covariates [2]. Logistic regression was employed to construct a baseline model with these covariates and advanced models incorporating additional ADC radiomic features. Model performance was assessed using area under the receiver operating characteristics curve (AUC) and compared using likelihood ratio test. Finally, in the radiogenomics cohort which included both T0 ADC images and gene expression profiles, differentially expressed genes and enriched HALLMARK pathways from Molecular Signatures Database associated with the identified prognostic radiomic features were investigated [3].

Results: A total of 7 representative radiomic features were identified for 7 feature clusters. Through univariate and multivariate analysis, T0 Maximum3DDiameter, ΔT1 DependenceNonUniformityNormalized, ΔT1 GrayLevelNonUniformity, ΔT1 LargeDependenceEmphasis, and ΔT2 Elongation were found to be independently associated with pCR. Advanced models incorporating these features demonstrated statistically significant improvements over the baseline model, with AUC increasing from 0.751 to 0.772-0.797 in the training set and from 0.739 to 0.767-0.794 in the testing set, except that ΔT2 Elongation model didn’t show improvement in the testing set. The identified prognostic radiomic features were uniquely correlated with immune-related pathway HALLMARK_COMPLEMENT, signal transduction pathway involved in cell proliferation HALLMARK_KRAS_SIGNALING, and steroid hormone signaling-related pathway HALLMARK_SPERMATOGENESIS, potentially suggesting their unique contributions to pCR prediction.