Automated segmentation of the hypothalamus and associated subunits in brain MRI

Abstract

Despite the crucial role of the hypothalamus in the regulation of the human body, neuroimaging studies of this structure and its nuclei are scarce. Such scarcity partially stems from the lack of automated segmentation tools, since manual delineation suffers from scalability and reproducibility issues. Due to the small size of the hypothalamus and the lack of image contrast in its vicinity, automated segmentation is difficult and has been long neglected by widespread neuroimaging packages like FreeSurfer or FSL. Nonetheless, recent advances in deep machine learning are enabling us to tackle difficult segmentation problems with high accuracy. In this paper we present a fully automated tool based on a deep convolutional neural network, for the segmentation of the whole hypothalamus and its subregions from T1-weighted MRI scans. We use aggressive data augmentation in order to make the model robust to T1-weighted MR scans from a wide array of different sources, without any need for preprocessing. We rigorously assess the performance of the presented tool through extensive analyses, including: inter- and intra-rater variability experiments between human observers; comparison of our tool with manual segmentation; comparison with an automated method based on multi-atlas segmentation; assessment of robustness by quality control analysis of a larger, heterogeneous dataset (ADNI); and indirect evaluation with a volumetric study performed on ADNI. The presented model outperforms multi-atlas segmentation scores as well as inter-rater accuracy level, and approaches intra-rater precision. Our method does not require any preprocessing and runs in less than a second on a GPU, and approximately 10 seconds on a CPU. The source code as well as the trained model are publicly available at https://github.com/BBillot/hypothalamus_seg, and will also be distributed with FreeSurfer.

Keywords: Convolutional neural network; Hypothalamus; Public software; Segmentation.

Fig. 1

Axial slices of intermediate image volumes obtained at different steps of the proposed augmentation model. First, the input image (a) is spatially deformed (b). We then apply a random bias field (c), and further global intensity augmentation (d). Finally the image is flipped along the right/left axis with a probability of 0.5. Each row corresponds to a different subject. The displayed slices correspond to the same coordinate in the inferior-superior axis. Augmentation is performed on the fly, and all random parameters are resampled at every step in training, such that the network is never exposed to the same image twice.

Fig. 2

Architecture of the 3D deep learning network. The first layer comprises 24 kernels, this number being doubled after each max-pooling, and halved after each up-convolution.

Fig. 3

Example of manually segmented hypothalamus in (a) sagittal, (b) axial and (c) coronal views. (d) 3D rendering of the right hypothalamus. Subunits are depicted in different colours: a-sHyp in blue, a-iHyp in yellow, supTub in green, infTub in pink, and posHyp in orange.

Fig. 4

Comparison between coronal slices of manual and automated segmentations for two subjects randomly selected from the internal dataset. Slices are shown from anterior (left) to posterior (right). The four rows associated to each subject respectively illustrate the original image, the manual ground truth (GT), the segmentation produced by MAS, and the segmentation of the proposed network. Subunits colours follow the same scheme as in Fig. 3.

Fig. 5

Comparison between MAS and our network on the test scans of the internal dataset: (a) Dice coefficients, (b) average boundary distance, and (c) Hausdorff distance. The improvement of the network is statistically significant for all metrics at the ${10}^{-3}$level (two-sided non-parametric Wilcoxon signed-rank tests) for the whole hypothalamus and all subunits. For each box, the central mark is the median; edges are the first and third quartiles; whiskers extend to 1.5 interquartile ranges around the median; and outliers are marked with ✦.

Fig. 6

Comparison between the intra-rater, inter-rater, and automated segmentations scores on the ten subjects from the variability experiment: (a) Dice coefficients, (b) average boundary distance (mm), and (c) Hausdorff distance (mm). Statistical significance (two-sided non-parametric Wilcoxon signed-rank test) is represented by black circles (•p < 0.05, ••p < 0.01).

Fig. 7

Coronal slices of segmentations produced by the network for four subjects randomly selected from the ADNI dataset. The first two cases are drawn among the control group, and the other two among AD subjects. Slices are shown from anterior (left) to posterior (right).

Fig. 8

Coronal slices of segmentations produced by the network for two subjects of the external dataset (one from each subdataset).