Infusing Domain Knowledge into nnU-Nets for Segmenting Brain Tumors in MRI

Accurate and reproducible segmentation of brain tumors from multi-modal magnetic resonance (MR) scans is a pivotal step in clinical 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 deep models 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. Our approach was also validated over the hold-out testing data which encompassed the BraTS 2021 Challenge test set, as well as new data from out-of-sample sources including independent pediatric population of diffuse intrinsic pontine glioma patients, together with an independent multi-institutional dataset covering under-represented Sub-Saharian African adult patient population of brain diffuse glioma. Our technique was ranked 2 ^nd and 3 ^rd over the pediatric and Sub-Saharian African populations, respectively, proving its high generalization capabilities.