Expert Model Prediction Through Feature Matching

Supervised brain MRI segmentation performance relies on test sample alignment to the training domain. This is a function of various factors outside practical control such as imaging artefacts and demographics. One way of alleviating this risk in a automated segmentation pipeline is through a pre-segmentation domain alignment test. We explore a potential solution in the form of expert models created through clustering. We use the BraTS-2023 dataset to cluster into four groups reflecting medical consensus followed by baseline specialisation. We find that while the expert performance does not significantly outperform the baseline, the ensemble of these experts does. To scrutinise the results further we examine the performance on tumour growth segmentation of the various methods and find that the non-ensemble experts perform the best in this regard. Finally, we propose an independent performance indicator which may be used to inform aleatoric uncertainty estimation. Code available at: https://github.com/bip5/ExpertModels .