Knowing what you know in brain segmentation using Bayesian deep neural networks

CEREBRUM: a fast and fully-volumetric Convolutional Encoder-decodeR for weakly-supervised sEgmentation of BRain strUctures from out-of-the-scanner MRI

“…QuickNAT [12] leverages a 2D based approach to efficiently segment brain MRI, exploiting a paradigm that aggregates the predictions of three different encoder-decoder models by averaging the probability maps -each model trained to segment a single slice at a time along one of the three principal axes (longitudinal, sagittal, and coronal). MeshNet [14,16] is a feedforward CNN based on 3D dilated convolutions, whose structure guarantees good results while keeping the number of parameters low. NeuroNet [11] is an encoder-multi-decoder CNN, trained to replicate segmentation results obtained with multiple state-of-the-art neuroimaging tools.…”

“…However, none of these studies provide a manually annotated ground-truth (GT), as carrying out the operation on such large databases would prove exceptionally time-consuming. For this reason, most of the studies investigating the application of Deep Learning architectures for brain MRI segmentation make use of automatically produced GT for training purposes [12,16,14,11] -with some of them reporting the latter can be exploited to train models that perform the same [11], or even better [12], than the automated pipeline itself. Motivated by this rationale, to train and test the proposed model we exploit a large collection of out-of-the-scanner MR images and the results of the FreeSurfer [20] cortical reconstruction process recon-all as reference GT.…”

“…Although this method simplifies the process from a technical point of view, it introduces significant limitations in the analysis: since each 2D or 3D patch is segmented independently from the others, these models mostly exploit local spatial information -ignoring "global" cues, such as the absolute and relative positions of different brain structures -which makes them sub-optimal. Different works have considered the potential improvements of removing said volume partitioning [16,13]. Such fully-volumetric approach has already been applied to prostate [17], heart atrium [18], and proximal femur MRI segmentation [19], but not yet in the context of brain MRI segmentationwhere it could prove particularly useful given the complex geometry and the variety of structures characterising the brain anatomy.…”