Federated Evaluation of nnU-Nets Enhanced with Domain Knowledge for Brain Tumor Segmentation

Accurate and reproducible segmentation of brain tumors from multi-modal magnetic resonance (MR) scans is a pivotal step in practice. In this BraTS Continuous Evaluation initiative, we exploit a 3D nnU-Net for this task which was ranked at the 6^th place (out of 1600 participants) in the BraTS’21 Challenge. We benefit from an ensemble of deep models enhanced with the expert knowledge of a senior radiologist captured in a form of several post-processing routines. The experimental study showed that infusing the domain knowledge into the algorithm can enhance their performance, and we obtained the average Dice score of 0.81977 (enhancing tumor), 0.87837 (tumor core), and 0.92723 (whole tumor) over the validation set. For the test data, we had the average Dice score of 0.86317, 0.87987, and 0.92838 for the enhancing tumor, tumor core and whole tumor. To validate the generalization capabilities of the nnU-Nets enhanced with domain knowledge, we performed their federated evaluation within the Federated Tumor Segmentation (FeTS) 2022 Challenge over the datasets captured across 30 institutions. Our technique was ranked 2^nd across all participating teams, proving its generalization capabilities over unseen out-of-sample datasets.