Researchers use information theory to cut MRI inputs by half without losing accuracy
Choosing just two MRI sequences achieves 98.5% of full four-sequence segmentation accuracy.
A new paper from Agamdeep Chopra and Mehmet Kurt introduces a Partial Information Decomposition (PID) framework that ranks MRI contrast pairs by their redundant, unique, and synergistic information about regional tumor burden. The goal: identify which two of the four standard sequences (T1n, T1c, T2w, T2-FLAIR) provide the most useful information for 3D brain tumor segmentation, enabling resource-constrained deep neural network training. The framework selected T1c+T2-FLAIR as the highest-ranked pair, outperforming all other two-input combinations.
Eleven architecturally identical lightweight 3D U-Nets were trained on different input configurations. On an independent test cohort, the T1c+T2-FLAIR model achieved a mean Dice score of 0.676, while the full four-input model scored 0.687—a difference of just 1.6%. Independent Shapley analysis on the full-input model confirmed T2-FLAIR and T1c as the most influential inputs, with their pairwise interaction being the strongest. These results validate PID as a practical pre-training selection strategy, allowing researchers to cut data and computational demands in half without meaningful accuracy loss, critical for clinical deployment on limited hardware.
- PID framework ranked all 6 possible MRI contrast pairs, selecting T1c+T2-FLAIR as optimal for brain tumor segmentation.
- Lightweight 3D U-Nets using only two sequences achieved mean Dice 0.676 vs 0.687 for all four sequences (1.6% drop).
- Shapley analysis independently confirmed T2-FLAIR and T1c as the most influential inputs, validating PID's selection.
Why It Matters
Enables accurate brain tumor segmentation with half the MRI data, reducing computational costs for resource-limited clinical settings.