CerebNet: A fast and reliable deep-learning pipeline for detailed cerebellum sub-segmentation

Abstract

Quantifying the volume of the cerebellum and its lobes is of profound interest in various neurodegenerative and acquired diseases. Especially for the most common spinocerebellar ataxias (SCA), for which the first antisense oligonculeotide-base gene silencing trial has recently started, there is an urgent need for quantitative, sensitive imaging markers at pre-symptomatic stages for stratification and treatment assessment. This work introduces CerebNet, a fully automated, extensively validated, deep learning method for the lobular segmentation of the cerebellum, including the separation of gray and white matter. For training, validation, and testing, T1-weighted images from 30 participants were manually annotated into cerebellar lobules and vermal sub-segments, as well as cerebellar white matter. CerebNet combines FastSurferCNN, a UNet-based 2.5D segmentation network, with extensive data augmentation, e.g. realistic non-linear deformations to increase the anatomical variety, eliminating additional preprocessing steps, such as spatial normalization or bias field correction. CerebNet demonstrates a high accuracy (on average 0.87 Dice and 1.742mm Robust Hausdorff Distance across all structures) outperforming state-of-the-art approaches. Furthermore, it shows high test-retest reliability (average ICC >0.97 on OASIS and Kirby) as well as high sensitivity to disease effects, including the pre-ataxic stage of spinocerebellar ataxia type 3 (SCA3). CerebNet is compatible with FreeSurfer and FastSurfer and can analyze a 3D volume within seconds on a consumer GPU in an end-to-end fashion, thus providing an efficient and validated solution for assessing cerebellum sub-structure volumes. We make CerebNet available as source-code (https://github.com/Deep-MI/FastSurfer).

Keywords: CerebNet; Cerebellum; Computational neuroimaging; Deep learning.

Fig. 1

Segmentation examples of a fully automated segmentation of CerebNet in a healthy control (A, B) as well as a symptomatic SCA3 patient (C, D) projected onto a coronal and sagittal slice.

Fig. 2

Dice score (larger values are better) and Robust Hausdorff Distances (HD95, smaller values are better) on validation cases for comparison of baseline, individual method contributions (not cumulative) and CerebNet. CerebNet combines multiple data augmentations with pre-training on a representative cross-study dataset. Deformation-250/500 indicates the number of realistic deformation fields used for static augmentation. The baseline model is our network without augmentation or pre-training. Error bars indicate 95% confidence intervals.

Fig. 3

Comparison of mean a) Dice score (larger values are better), b) Volume similarity (larger values are better), c) Robust Hausdorff Distance (HD95, smaller values are better), and d) Hausdorff Distance (HD, smaller values are better) over all structures and participants. CerebNet outperforms both ACAPULCO^rt (which is retrained on our dataset for direct comparison) and SUIT+FS. Error bars indicate 95% confidence intervals. Statistical significance for all results is confirmed by two-sided non-parametric Wilcoxon signed-rank tests ($**$: $p<.01$).

Fig. 4

Dice score (larger values are better) and Robust Hausdorff Distance (HD95) (smaller values are better) per sub-structure for CerebNet, ACAPULCO^rt and SUIT+FS. Illustrations show the cross-subject average of the metric (bar) and corresponding, bootstrapped 95% confidence intervals (error bars), data points (eight per bar, may overlap) as well as the significance level calculated by a Wilcoxon signed-rank test ($*:p<.05$ and $**:p<.01$). CGM: Cerebellar Gray Matter; CWM: Cerebellar White Matter.

Fig. 5

Comparison of Inter-rater reliability and CerebNet by Dice score and Robust Hausdorff Distance (HD95) per sub-structure. Error bars indicate 95% confidence intervals. CGM is Cerebellar Gray Matter and CWM is Cerebellar White Matter ($*:p<.05$ and $**:p<.01$).

Fig. 6

Intraclass correlation coefficient (ICC) on volume of Kirby and OASIS1 datasets for test-retest analysis. Error bars indicate the 95% confidence interval. Statistical significance is calculated with a two-sided non-parametric Wilcoxon signed-rank test over the absolute volume difference, since ICC values cannot provide significance information. $*$ and $**$ annotations represent statistical significance for better volume consistency with $p<.05$ and $p<.01$, respectively.

Fig. 7

Map of volume change in HC vs. pre-ataxic SCA3 (top) and pre-ataxic vs. ataxic SCA3 (bottom). Per-region p-values of the respective group comparisons are shown for 3 different methods: CerebNet, ACAPULCO (as distributed by Han et al., 2020b) and ACAPULCO^rt (ACAPULCO retrained on our dataset). Red colors indicate atrophy, blue colors indicate volume increase (color saturation corresponds to significance). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 8

Comparison of Dice and Robust Hausdorff Distance (HD95) metrics for CerebNet, the retrained ACAPULCO^rt, SUIT+FS as well as the original (published) ACAPULCO on our test set. Note, the direct comparison of CerebNet and the original ACAPULCO does not correct for the differences in the training datasets; ACAPULCO^rt corrects for this difference (see text).

Fig. 9

Qualitative “out-of-distribution” evaluation of CerebNet: Segmentation maps for pathologies, which are not part of the training (SCA1, SCA2, SCA6, MSA-C, RFC1, SYNE1, AOA2 and CTX) together with an in-distribution example (SCA3). Images illustrated here are randomly picked from a larger repository of images and represent average performance. *: Label is not visible in the shown slice.