IMG-09. A DEEP LEARNING-BASED APPROACH FOR BRAIN TISSUE EXTRACTION USING MULTI- AND SINGLE-PARAMETRIC MRI IN PEDIATRICS

Abstract BACKGROUND Skull-stripping, the process of extracting brain tissue from MR images, is an important step for tumor segmentation and downstream imaging-based analytics such as AI-powered radiomic feature extraction. Existing skull-stripping models, designed for pediatric or adult patients, show limitations in accurately segmenting tumors in sellar/suprasellar regions. This limitation hinders their reliable application across different histologies of pediatric brain tumors. We propose a deep learning approach for fully automated skull-stripping, compatible with both single- or multi-parametric MRI sequences. METHODS We developed 3D nnU-Net models trained on preprocessed MRI sequences (including pre- and post-contrast T1w, T2w, and FLAIR) from 336 patients with brain tumors across multiple tumor histologies such as low-grade, high-grade and brainstem gliomas, medulloblastoma, ependymoma, etc., aged between 3 months and 20 years (median age, 8.5 years). The training utilized manually generated brain masks, including the sellar/suprasellar region, from 153 patients and employed 5-fold cross-validation to split the data into inner training-validation sets. The models were then tested on a withheld set of 183 subjects. Additionally, we trained a single-parametric model on individual images, resulting in 612 training and 732 testing cases. Model performance was evaluated using the Dice similarity metric for segmenting both the entire brain and slices specifically containing the sella turcica. RESULTS The multi-parametric and single-parametric models achieved mean±sd Dice scores of 0.981±0.008 (median=0.983) and 0.979±0.009 (median=0.981), respectively. For the sellar/suprasellar slices, the scores were 0.983±0.009 (median=0.986) and 0.981±0.012 (median=0.984), respectively. These results indicate a high precision in segmenting not only the entire brain volume, but also the sellar/suprasellar region. CONCLUSION Our proposed deep learning-based skull-stripping approach, leveraging both multi-parametric and single-parametric MRI inputs, demonstrates excellent accuracy. These models, made publicly available, have potential for improving auto-processing pipelines in pediatric brain tumors.