MENINGIOMA SEGMENTATION USING DEEP LEARNING FOR CLINICAL PRACTICE

Abstract AIMS An estimated 34,840 meningioma cases will be diagnosed in the UK this year. Consistent segmentation across institutions and clinicians is crucial for volumetric assessment, growth monitoring, and radiotherapy planning. This study presents a deep learning-based automated segmentation approach to evaluate its clinical integration potential for improved workflows and patient care. METHODS We trained the nnU-Net architecture on the BraTS Meningioma 2023 dataset, comprising 1,000 multi-sequence MR images from six academic centres. The dataset includes T2-, T2/FLAIR-, T1-, and post-contrast T1-weighted sequences with expert-refined segmentations of three meningioma sub-compartments. RESULTS Using 5-fold cross-validation, the model achieved a lesion-wise mean DSC of 0.935, 0.922 and 0.924 for enhanc- ing tumour (ET), tumour core (TC), and whole tumour (WT), respectively. This performance surpassed existing literature (pooled DSC of 88%, range: 84-91%) of 5 studies with internal validation datasets. Although our per- formance exceeded the published challenge winner (0.899, 0.904, and 0.871 for ET, TC, and WT, respectively), 424 validation and hidden test cases, currently unavailable to us, were not included in the evaluation. Analysis of the lowest 5% of Dice scores, in collaboration with a consultant neuroradiologist, revealed reduced accuracy in segmenting lesions in the para-clinoid and cavernous sinus regions. However, the model exhibited increased sensitivity, detecting additional convexity meningiomas absent from the BraTS segmentations. CONCLUSION Our study demonstrates the feasibility of a highly accurate, fully automatic meningioma segmentation tool for clinical implementation. Future work will focus on improving the model performance and validating clinical applicability using a diverse, in-house dataset with varying patient populations and missing sequences.