2281 - Early MRI Radiomics-Based Triage Model Predicts Response and Survival after Bevacizumab in Recurrent Glioblastoma

03:00pm - 04:00pm ET

Poster Hall - Exhibit Hall A

Screen: 13

POSTER

Presenter(s)

Weir Chiang You, MD, PhD - Taichung Veterans General Hospital, Taichung, Taichung

W. C. You¹, Y. T. Shao², Y. C. Chen³, and Y. Y. Hsu⁴; ¹Department of Post-Baccalaureate Medicine, National Chung Hsing University, Taichug City, Taiwan, ²Department of Post-Baccalaureate Medicine, National Chung Hsing University, Taichung City, Taiwan, ³Department of Computer Science and Engineering, National Chung Hsing University, Taichung City, Taiwan, ⁴Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung City, Taiwan

Purpose/Objective(s): Radiomics has emerged as a promising tool in neuro-oncology; however, the temporal stability and transferability of imaging-derived signatures across the disease trajectory of glioblastoma remain unclear. We investigated whether a machine-learning radiomics signature trained exclusively on preoperative MRI to predict short-term morphologic response could generalize to prognostic stratification in an independent cohort of recurrent glioblastoma patients undergoing bevacizumab therapy.

Materials/Methods: A leakage-controlled machine-learning pipeline incorporating recursive feature elimination with random forest (RF-RFE) and nested cross-validation was developed. A radiomics signature (cascade random forest classifier) was trained using contrast-enhanced T1-weighted preoperative MRI from 122 primary glioblastoma patients to predict short-term objective response rate (ORR). Without retraining, temporal transferability and prognostic performance were evaluated in an independent external validation cohort (N = 19) consisting of pre-bevacizumab MRIs from recurrent GBM patients. Performance was compared with conventional radiologic assessment for ORR prediction (area under the curve [AUC], accuracy) and for survival stratification using Kaplan–Meier analysis and log-rank testing for post-bevacizumab progression-free survival and overall survival (AFT and AOST).

Results: The RF-RFE pipeline identified 10 core radiomic features predominantly reflecting intratumoral heterogeneity, including gray level size zone matrix–derived metrics. In the recurrent cohort, conventional radiologic assessment demonstrated strong short-term ORR prediction (AUC = 0.89; accuracy = 89.5%), whereas the radiomics signature showed moderate discrimination (AUC = 0.64; accuracy = 57.9%). However, survival analysis revealed discordant findings: conventional ORR assessment failed to stratify post-bevacizumab survival (AOST: log-rank p = 0.236; AFT: p = 0.358). In contrast, the preoperative radiomics signature significantly stratified survival outcomes, with AI-predicted responders exhibiting prolonged AOST (p = 0.015) and AFT (p = 0.014).

Conclusion:

A radiomics signature derived from baseline preoperative MRI demonstrated temporal transferability to the recurrent, pre-bevacizumab setting. Although conventional radiologic assessment accurately captures short-term radiographic response, it does not reliably translate into survival benefit, potentially due to pseudoresponse phenomena. Conversely, the radiomics signature appears to capture intrinsic tumor phenotypes associated with long-term outcomes. Baseline radiomics may therefore serve as a prognostic biomarker for post-bevacizumab survival in glioblastoma.