An automatic multi-tissue human fetal brain segmentation benchmark using the Fetal Tissue Annotation Dataset

Abstract

It is critical to quantitatively analyse the developing human fetal brain in order to fully understand neurodevelopment in both normal fetuses and those with congenital disorders. To facilitate this analysis, automatic multi-tissue fetal brain segmentation algorithms are needed, which in turn requires open datasets of segmented fetal brains. Here we introduce a publicly available dataset of 50 manually segmented pathological and non-pathological fetal magnetic resonance brain volume reconstructions across a range of gestational ages (20 to 33 weeks) into 7 different tissue categories (external cerebrospinal fluid, grey matter, white matter, ventricles, cerebellum, deep grey matter, brainstem/spinal cord). In addition, we quantitatively evaluate the accuracy of several automatic multi-tissue segmentation algorithms of the developing human fetal brain. Four research groups participated, submitting a total of 10 algorithms, demonstrating the benefits the dataset for the development of automatic algorithms.

Fig. 1

Left: Examples of fetal brain SR volumes with different quality ratings; Right: Quality ratings overview (1: poor quality; 2: good quality; 3: excellent quality).

Fig. 2

Example of Manual Segmentation (Dark Green: external cerebrospinal fluid; yellow: GM; brown: WM; blue: ventricles; bright red: cerebellum; light red: deep GM: bright green: brainstem/spinal cord).

Fig. 3

Analysis of the 3 annotator segmentations of all 9 volumes averaged together, and split into the categories of normal SR volumes, pathological SR volumes, excellent quality SR volumes, good quality SR volumes, and poor quality SR volumes.

Fig. 4

Number of atlas candidates (up to 2 GA difference and acceptable SR quality) vs number of atlases (5 best ranked NCC and NCC ≥ 0.8) used in the label fusion step, for the training dataset, using a leave-one-out approach.

Fig. 5

Overview of the metrics (DSC, HD95, VS) of all subjects for each algorithm (all tissue labels combined), and the corresponding algorithm ranking.

Fig. 6

An overview of the algorithms (all labels) evaluated on top: excellent quality SR volumes; middle: good quality SR volumes; and bottom: poor quality SR volumes, as well as the ranking for each of the quality levels. The poor-quality SR volumes have worse metrics (lower DSC and VS, higher HD95) than the average and excellent quality SR volumes, and the standard deviations are larger for every label. The difference in metrics between the average and excellent quality metrics is not noticeable.

Fig. 7

An overview of algorithms (all labels) evaluated on (a) Pathological SR volumes and (b) Non-pathological SR volumes, as well as the ranking for each. The segmentations of the normal SR volumes scored much higher than the pathological segmentations.