Training data requirements for atlas-based brain fibre tract identification

Large volumes of annotated training data are often required for data-driven image analysis methods. We consider two techniques for identifying brain fibre bundles from diffusion MRI scans, tractfinder and TractSeg, and compare performances using different amounts of training data. Our results show that tractfinder, an atlas-based method, shows no improvement in performance beyond a relatively small number of training samples. This is an advantage in a field where generating and maintaining high quality reference data is difficult and time-consuming