Abstract
Multi-contrast 3D MRI segmentation can be computationally demanding when all available sequences are used. We evaluate a pre-training Partial Information Decomposition framework that ranks input pairs based on their redundant, unique, and synergistic information regarding regional tumor burden and selects the highest-ranked pair for downstream training. This framework was applied to T1n, T1c, T2w, and T2-FLAIR MRI, resulting in the selection of T1c+T2-FLAIR.
We then trained eleven architecturally identical lightweight 3D U-Nets using different input configurations. On an independent test cohort, T1c+T2-FLAIR proved to be the strongest two-input configuration, achieving a Dice score of 0.676 compared to 0.687 for all four inputs, ranking second overall. Independent Shapley analysis on the full-input model identified T2-FLAIR and T1c as the most influential inputs, with their pairwise interaction being the strongest. These findings highlight the practical value of PID-based pre-training selection for identifying compact, informative MRI input sets before costly 3D model development.
Blogger's Review: This study introduces an effective MRI input selection strategy that optimizes deep learning training for brain tumor segmentation through Partial Information Decomposition, reducing resource consumption. This approach not only enhances model performance but also provides strong technical support for future research. It will be interesting to see how this framework can be further expanded to other medical imaging domains.