P04.18.A GLIOSEG: AN INTEGRATED FRAMEWORK FOR ROBUST AI-BASED GLIOMA SEGMENTATION THROUGH MODEL ENSEMBLING

Abstract BACKGROUND Treatment planning and surveillance of glioma patients benefit from accurate tumor segmentation. However, manual segmentation is time-consuming and subject to intra- and inter-rater variability. Recent advances in artificial intelligence (AI) have led to automated methods with performance comparable to expert human raters. Nevertheless, many tools rely on specific preprocessing steps or consist of separate installable components, which complicates their usage in research settings. We present Glioseg, an automated pipeline that delineates glioma subregions (edematous/infiltrated tissue, necrotic core, enhancing tumor) using multiple pre-trained AI models. The framework has been designed to be easily deployable by minimizing required preprocessing and usage configurations. MATERIAL AND METHODS Glioseg accepts pre-operative T1-weighted, contrast-enhanced T1-weighted, T2-weighted, and T2-weighted-FLAIR MRI. Preprocessing steps include reorientation to standard space, resampling, registration, and skull-stripping. During registration, patient scans are spatially aligned to each other, followed by alignment to a brain atlas. Segmentation is performed leveraging five pre-trained AI models, whose outputs are postprocessed and ensembled through simultaneous truth and performance level estimation (STAPLE). Final segmentations are transformed back to the native image space to facilitate downstream analysis. RESULTS Glioseg was evaluated on an external dataset of 43 patients with adult-type glioma of varying WHO CNS5 (2021) subtypes and grades. Generated segmentations were compared in native space to a semi-automatic ground truth, where outputs from an external AI model were corrected by a group of radiologists. Glioseg achieved an average whole-tumor Dice score of 0.85 ± 0.09 and multi-class Dice score of 0.77 ± 0.151. The lower value for the latter reflects the increased difficulty of segmenting individual compartments, particularly for smaller lesions. In comparison, the five individual segmentation models achieved average whole-tumor Dice scores of 0.85/0.86/0.79/0.84/0.78 and multi-class Dice scores of 0.70/0.70/0.74/0.50/0.70, demonstrating the added value of model ensembling. CONCLUSION Glioseg provides an automated solution for multi-compartment glioma segmentation, integrating pre-trained AI with standardized pre- and postprocessing and model ensembling. Initial validation demonstrates robust performance on a heterogeneous dataset. Ongoing work includes comparison to fully manual expert segmentations, benchmarking against established methods in the literature, and refinement of postprocessing to improve segmentation quality on smaller lesions. With its robustness and ease of use, Glioseg has the potential to become a standard tool to support glioma research through accurate and consistent tumor segmentations.