AUTHOR=Huang Jacky , Yagmurlu Banu , Molleti Powell , Lee Richard , VanderPloeg Abigail , Noor Humaira , Bareja Rohan , Li Yiheng , Iv Michael , Itakura Haruka 

TITLE=Brain tumor segmentation using deep learning: high performance with minimized MRI data

JOURNAL=Frontiers in Radiology

VOLUME=Volume 5 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/radiology/articles/10.3389/fradi.2025.1616293

DOI=10.3389/fradi.2025.1616293

ISSN=2673-8740

ABSTRACT=PurposeBrain tumor segmentation with MRI is a challenging task, traditionally relying on manual delineation of regions-of-interest across multiple imaging sequences. However, this data-intensive approach is time-consuming. We aimed to optimize the process by using a deep learning (DL) based model while minimizing the number of MRI sequences required to segment gliomas.MethodsWe trained a 3D U-Net DL model using the annotated 2018 MICCAI BraTS dataset (training dataset, n = 285), focusing on sub-segmenting enhancing tumor (ET) and tumor core (TC). We compared the performances of models trained on four different combinations of MRI sequences: T1C-only, FLAIR-only, T1C + FLAIR and T1 + T2 + T1C + FLAIR to evaluate whether a smaller MRI data subset could achieve comparable performance. We evaluated the performance on the four different sequence combinations using 5-fold cross-validation on the training dataset, then on our test dataset (n = 358) consisting of samples from a separately held-out 2018 BraTS validation set (n = 66) and 2021 BraTS datasets (n = 292). Dice scores on both cross-validation and test datasets were assessed to measure model performance.ResultsDice scores on cross-validation showed that T1C + FLAIR (ET: 0.814, TC: 0.856) matched or outperformed those of T1 + T2 + T1C + FLAIR (ET: 0.785, TC: 0.841), T1C-only (ET: 0.781, TC: 0.852) and FLAIR-only (ET: 0.008, TC: 0.619). Results on the test dataset also showed that T1C + FLAIR (ET: 0.867, TC: 0.926) matched or outperformed those of T1 + T2 + T1C + FLAIR (ET: 0.835, TC: 0.908), T1C-only (ET: 0.726, TC: 0.928), and FLAIR-only (ET: 0.056, TC: 0.543). T1C + FLAIR excelled in both ET and TC, exceeding the performance of the four-sequence dataset. T1C-only matched T1C + FLAIR in TC performance. Similarly, T1C and T1C + FLAIR also outperformed in ET delineation by sensitivity (0.829) and Hausdorff distance (5.964) on the test set. Across all configurations, specificity remained high (≥0.958). T1C performed well in TC delineation (sensitivity: 0.737), but the inclusion of all sequences led to improvement (0.754). Hausdorff distances clustered in a narrow range (17.622–33.812) for TC delineation across the configurations.ConclusionsDL-based brain tumor segmentation can achieve high accuracy using only two MRI sequences (T1C + FLAIR). Reduction of multiple sequence dependency may enhance DL generalizability and dissemination in both clinical and research contexts. Our findings may ultimately help mitigate human labor intensity of a complex task integral to medical imaging analysis.