P04.17.B DEEP LEARNING-BASED POST-OPERATIVE GLIOBLASTOMA SEGMENTATION AND TREATMENT RESPONSE ASSESSMENT

Abstract BACKGROUND Glioblastoma is a highly aggressive brain tumor with a poor prognosis. Accurate assessment of treatment response is critical for clinical decision-making with RANO 2.0 criteria being the reference standard. However, RANO 2.0 continues to emphasise manual bidimensional measurements, which are time-consuming and subjective. Existing automated response assessment methods provide a systematic approach to measurement, but often misclassify tumour boundaries leading to discrepancies when compared to expert RANO-based evaluations. The aim of the study was to use an automated tool to conduct RANO 2.0 response assessment, and optimise bidiemensional product measurement through post-processing methodologies. MATERIAL AND METHODS We developed a novel, fully automated end-to-end pipeline that follows RANO 2.0 criteria. The pipeline is based on the nnU-Net architecture and begins with standardised preprocessing of longitudinal postoperative MRI scans from a dataset collected at King’s College Hospital, UK. A BraTS 2024 pre-trained model was used to segment key tumour subregions, including the contrast-enhancing tumor, peritumoral oedema, and surgical cavity. The downstream assessment module implements tailored strategies to automatically compute the maximum bidimensional product within the contrast-enhancing region, whilst excluding surgical cavities to ensure accurate RANO-based evaluation. The reference standard for bidimensional glioma measurement was double expert radiologist measurement. Ablation studies were performed to assess the impact of the key components of tailored strategies, including slice selection and diameter measurement modules. RESULTS An evalution of automated hole-filling strategies showed that ignoring intra-enhancing region holes smaller than 5 mm² whilst assigning the bidimensional product to enhancing tissue, significantly improved measurement consistency between the reference standard and the automated measurement. The final model demonstrated strong agreement with expert annotations in bidimensional measurement. CONCLUSION The proposed automated pipeline replicates the manual RANO evaluation process more accurately when automated hole-filling strategies are employed. In the future, we aim to integrate the whole filling strategy as an option in RANO2.0-assist, an automated tool recently developed in our group, to offer a clinically deployable tool for fast and reliable glioblastoma response evaluation.