Brain metastases require accurate delineation for clinical treatment planning, yet magnetic resonance imaging acquisition protocols vary substantially across routine healthcare practice. In a retrospective study by Dong-Won Shin, Chang Hyun Lee, Hyun Ju Kim, Kihoon Sung, Young Jae Kim, Kwang Gi Kim, and Gi Taek Yee, the authors evaluated how heterogeneous T1-weighted MRI protocols impact lesion segmentation performance.
The researchers analyzed a retrospective imaging dataset of 925 contrast-enhanced T1-weighted MRI studies acquired from 720 patients. To evaluate protocol robustness, the team compared two three-dimensional architectures, 3D Attention U-Net and DUCK-Net 3D, executing five-fold cross-validation across the sequence configurations.
Models were trained under six sequence configurations, including multi-sequence, Black-Blood, Non-Black-Blood, and three single-sequence cohorts. Independent test datasets comprised three MRI cohorts: MPR Range, MPR Thick Range, and T1 SPACE 3D. Performance was evaluated using the Dice similarity coefficient, volumetric regression, and false discovery rate-corrected statistical comparisons.
DUCK-Net 3D demonstrated more consistent segmentation performance across the independent MRI cohorts, particularly under challenging cross-sequence conditions. Multi-sequence training achieved the most stable cross-sequence performance overall, whereas single-sequence models showed greater sensitivity and performance vulnerability to protocol variability.
Both deep learning architectures exhibited a noticeable decline in segmentation accuracy for small lesions measuring less than 0.4 milliliters.
Subject of Research: Impact of MRI sequence heterogeneity on deep learning segmentation of brain metastases
Article Title: Impact of MRI sequence heterogeneity on deep learning–based segmentation of brain metastases
Article References: Shin, D.-W., Lee, C. H., Kim, H. J., Sung, K., Kim, Y. J., Kim, K. G., & Yee, G. T. (2026). Impact of MRI sequence heterogeneity on deep learning–based segmentation of brain metastases. PLOS One, 21(11), e0363055. https://doi.org/10.1371/journal.pone.0363055
Image Credits: AI Generated
DOI: 10.1371/journal.pone.0363055
Keywords: brain metastases, magnetic resonance imaging, protocol heterogeneity, deep learning segmentation, 3D Attention U-Net, DUCK-Net 3D, T1-weighted MRI, Black-Blood sequence, Non-Black-Blood sequence, MPR Range, MPR Thick Range, T1 SPACE 3D, Dice similarity coefficient, volumetric regression, cross-sequence robustness, small lesion segmentation, treatment planning
Cite Scienmag News
Cassandra Pierce. (November 21, 2026). Diverse MRI Training Improves Deep Learning Brain Metastasis Segmentation Across Protocols. Scienmag. https://scienmag.com/diverse-mri-training-improves-deep-learning-brain-metastasis-segmentation-across-protocols/
Cassandra Pierce. "Diverse MRI Training Improves Deep Learning Brain Metastasis Segmentation Across Protocols." Scienmag, 21 November 2026, https://scienmag.com/diverse-mri-training-improves-deep-learning-brain-metastasis-segmentation-across-protocols/. Accessed 21 November 2026.
Cassandra Pierce. "Diverse MRI Training Improves Deep Learning Brain Metastasis Segmentation Across Protocols." Scienmag. November 21, 2026. https://scienmag.com/diverse-mri-training-improves-deep-learning-brain-metastasis-segmentation-across-protocols/








