PID Optimizes MRI Selection for Brain Tumor Segmentation

This research explores using a Partial Information Decomposition (PID) framework to optimize the selection of 3D MRI sequences for brain tumor segmentation, particularly in resource-constrained deep neural network training environments. The goal is to identify the most informative subset of MRI contrasts, such as T1n, T1c, T2w, and T2-FLAIR, to reduce computational load without significantly sacrificing accuracy. The PID framework ranks input pairs based on their redundant, unique, and synergistic information content regarding tumor burden. Applying this framework, T1c+T2-FLAIR was identified as the highest-ranked two-input combination. Subsequent training of lightweight 3D U-Nets confirmed that this pair performed nearly as well as using all four inputs, achieving a mean Dice score of 0.676 compared to 0.687 for the full set. Independent Shapley analysis further corroborated these findings, highlighting T2-FLAIR and T1c as the most influential inputs. This demonstrates PID's practical value in selecting compact, yet highly informative, MRI input sets for efficient 3D model development.