Quality-Conditioned Region-Specific Learning for Multimodal 3D Brain Tumor Segmentation

Abstract

Accurate segmentation of brain tumors from multimodal Magnetic Resonance Imaging (MRI) is essential for quantitative tumor assessment and subsequent clinical analysis. However, segmentation of the enhancing tumor (ET) region remains particularly challenging because of its relatively small and heterogeneous appearance and its sensitivity to variations in MRI image quality. Although existing deep learning approaches have achieved substantial progress in multimodal brain tumor segmentation, most methods primarily focus on improving network architectures or loss functions, while the influence of modality-specific MRI quality on region-specific segmentation learning remains insufficiently addressed. In particular, the integration of quantitative MRI quality information with explicit learning of the distinct whole tumor (WT), tumor core (TC), and ET regions has received limited investigation. To address this gap, this study proposes a Quality-Conditioned Region-Specific Learning (QCR) framework for multimodal 3D brain tumor segmentation. The framework incorporates quantitative MRI quality characteristics into the segmentation network and explicitly conditions the learning process on WT, TC, and ET regions. Four MRI modalities, T1, T1ce, T2, and FLAIR, from the BraTS 2020 dataset were used. A 16-dimensional quality representation was constructed using coefficient of variation (CVP), signal-to-noise ratio measures (SNR3 and SNR4), and coefficient of joint variation (CJV) across the four modalities. The experiments included 368 valid subjects, divided into 257 training, 55 validation, and 56 independent testing subjects. The proposed QCR framework was implemented using a controlled 3D U-Net with Dice and weighted cross-entropy loss. On the independent test set, QCR achieved mean Dice scores of 0.8888 ± 0.1067 for WT, 0.7973 ± 0.2500 for TC, and 0.7331 ± 0.2360 for ET. Notably, the ET Dice improved from 0.7075 in the controlled baseline to 0.7331, representing an absolute improvement of 0.0256. The corresponding ET IoU, precision, and sensitivity were 0.6222 ± 0.2427, 0.7608 ± 0.2259, and 0.8084 ± 0.2216, respectively. An ablation study further showed that QCR produced comparable region-specific Dice performance to fixed region conditioning, while providing a quality-conditioned formulation that incorporates MRI quality characteristics into the learning process. These findings demonstrate the potential of combining quantitative MRI quality information with region-specific learning, particularly for improving segmentation of the challenging ET region.