3D assessment of fetal brain development in the second trimester from ultrasound volumes

1996 

Abstract Submission 

Madeleine Wyburd¹, INTERGROWTH-21st Consortium², Mark Jenkinson^3,4,5, Ana Namburete^1,3

¹Department of Computer Science, University of Oxford, Oxford, UK, ²Oxford Maternal & Perinatal Health Institute (OMPHI), Green Templeton College, University of Oxford, Oxford, United Kingdom, ³Wellcome Centre for Integrative Neuroimaging, FMRIB, Oxford, United Kingdom, ⁴Australian Institute for Machine Learning (AIML), Adelaide, Australia, ⁵South Australian Health and Medical Research Institute (SAHMRI), Adelaide, Australia

Madeleine Wyburd  
Department of Computer Science, University of Oxford
Oxford, UK

INTERGROWTH-21st Consortium  
Oxford Maternal & Perinatal Health Institute (OMPHI), Green Templeton College, University of Oxford
Oxford, United Kingdom

Mark Jenkinson  
Wellcome Centre for Integrative Neuroimaging, FMRIB|Australian Institute for Machine Learning (AIML)|South Australian Health and Medical Research Institute (SAHMRI)
Oxford, United Kingdom|Adelaide, Australia|Adelaide, Australia

Ana Namburete, Professor  
Department of Computer Science, University of Oxford|Wellcome Centre for Integrative Neuroimaging, FMRIB
Oxford, UK|Oxford, United Kingdom

Recent advances in fetal MRI have allowed the complex 3D growth of the developing cortical plate (CP) to be characterised [1-4]. This advance has improved our understanding of sex differences [2], asymmetrical development [2-3], and differences between healthy and at-risk pregnancies [4]. However, these studies are often limited by small sample sizes and a focus on the third trimester of pregnancy. In contrast, large ultrasound (US) datasets throughout pregnancy are obtainable as US has a short acquisition time and is routinely used across the world in clinical practice unlike MRI [5]. However, current methods for studying the 3D growth of the CP are limited to MRI and are not directly transferable to US due to significant domain differences. In this study, we propose an automated deep-learning pipeline to characterise the CP from 3D US scans, as shown in Fig. 1.

To measure the CP's properties, it must first be delineated. In US, large shadows often obstruct regions of the CP, leading to holes within the extracted surface that make further analysis extremely difficult. To overcome this, we used a deep learning-based topology-preserving segmentation network: TEDS-Net [6]. TEDS-Net (gθ) learns a deformation field (φ) that deforms a prior shape (P) to produce a segmentation (Ý), anatomically guiding the segmentation in the regions of shadows. For this task, we used a CP label from an US atlas [7] as P, to minimise the requirement for complex deformations. To train and evaluate the network, we used n=643 transabdominal US volumes between 18 and 26 weeks' gestation collected as part of the INTERGROWTH-21st (IG21) Project [5]. The trained network was then applied to a further n=2,188 unseen, unlabelled, IG21 volumes, and both global and local properties were measured from the predicted topologically-correct CP surfaces.

Our network achieved topologically correct segmentations with a Dice overlap of 80% compared to the manual CP labels. The segmentation performance was found to decrease significantly with increasing gestational age (Pearson correlation test: p<0.01); this is likely due to the greater calcification of the skull, which increases US artefacts making the CP boundaries less well defined. Volume (range = 5 - 55 cm3), surface area (range = 40-150 cm2) and 3D Sylvian Fissure (SF) depth (range = 5-12 mm) measures were consistent with previous MRI studies [3,8] and a 2D US study [9], as shown in Fig. 2A. As the chosen P was based on an atlas, it could be parcellated into five lobes (frontal, parietal, occipital, temporal and insula) by aligning P to a fetal MRI parcellation map [1]. Using φ, the parcellation map was registered to each individual's CP segmentation generating a personalised parcellation map (L), enabling regional measures of local features. The average cortical depth (D), thickness (T) and volume for each lobe are shown in Fig. 2B. The insula deepened and thickened at the fastest rate, which is expected due to SF opercularisation. The volumes of the frontal, parietal, and temporal lobes increased most across this gestational period, closely aligning with previous MRI growth curves [10]. As the transformation fields, φ, are invertible, the CP local properties, e.g. thickness and cortical depth, can be propagated back onto P and averaged across the population, facilitating a direct comparison between individuals and gestational weeks, as shown in Fig 2C.

In summary, we have developed the first automated pipeline for extracting and analysing the 3D CP from challenging US scans, taking less than 10s per scan. Our chosen approach enables efficient volume parcellation and groupwise comparisons, and the findings were consistent with previous MRI and US studies. Furthermore, this pipeline demonstrates that in-depth neurodevelopmental studies can now be conducted using US, potentially paving the way for more research in this modality in the future.

Early life, Adolescence, Aging

Normal Brain Development: Fetus to Adolescence ²

Methods Development

Segmentation and Parcellation ¹

Cortex

Development

Machine Learning

Structures

ULTRASOUND

Other - Cortical plate