Skip to Main content Skip to Navigation

Reconstruction 3D de données échographiques du cerveau du prématuré et segmentation des ventricules cérébraux et thalami par apprentissage supervisé

Abstract : About 15 million children are born prematurely each year worldwide. These patients are likely to suffer from brain abnormalities that can cause neurodevelopmental disorders : cerebral palsy, deafness, blindness, intellectual development delay, ...The volume of the brain structures is a clinical parameter that can be used to predict these disorders and to guide patients into appropriate health care pathways. In the case of the cerebral ventricular system (CVS), the volume is also used to determine when surgery should be performed. Today, these quantitative measurements can only be obtained by analyzing MRI data, which is an examination performed in only 15 % of premature infants. In clinical routine, 2D transfontanellar ultrasound (ETF) is performed on all premature infants.This examination is used to diagnose ventricular dilation but not to quantify precisely the ventricular volume or that of other brain structures because the 3D information is unavailable. The aim of this thesis is to show that, provided that the image quality is sufficient, 3D ETF would make it possible to acquire data in which the volume of brain structures could be quantified in 100 % of premature infants. The main issues associated with this objective are to obtain high quality 3D ultrasound data and to label the millions of voxels they contain in a clinical time (a few seconds). This thesis focuses on the segmentation of CVS and thalami. The four main contributions of this work are : the development of an algorithm that enables the high-quality 3D reconstruction of 2D ETFs (1), the creation of annotated 3D databases of the CVS and thalami (2), the segmentation of the CVS and the thalami in a clinical time by convolutional neural networks (CNN) (3) and finally the beginning of the creation of a CNN architecture dedicated to this segmentation problem that learns the anatomical position of CVS (4). Our reconstruction algorithm was used to reconstruct 25 high-quality ultrasound volumes. It was validated in-vivo where an accuracy of 0.69 $\pm$ 0.14 mm was obtained on the corpus callosum. First validation attempts were also performed on a neonate brain phantom. In total, 25 reference 3D segmentations were obtained for the CVS and 16 for the thalami. The best segmentation results were obtained with the V-net, a 3D CNN,which segmented the CVS and the thalami with respective Dice of 0.828 $\pm$ 0.044 et 0.891 $\pm$ 0.016. These segmentations were performed in a few seconds in volumes of size 320 $\times$ 320 $\times$ 320 voxels. Learning the anatomical position of the CVS was achieved by integrating a CPPN (Compositional Pattern Producing Network) into the CNNs. It significantly improved the accuracy of CNNs when they had few layers. For example, in the case of the 7-layer V-net network, the Dice has increased from 0.524 $\pm$ 0.076 to 0.724 $\pm$ 0.107. This thesis shows that it is possible to automatically segment brain structures of the premature infant into 3D ultrasound data with precision and in a clinical time. This proves that high quality 3D ultrasound could be used in clinical routine to quantify the volume of brain structures and paves the way for studies to evaluate its benefit to patients.
Complete list of metadatas

Cited literature [156 references]  Display  Hide  Download
Contributor : Matthieu Martin <>
Submitted on : Friday, February 21, 2020 - 4:07:52 PM
Last modification on : Saturday, February 22, 2020 - 1:39:49 AM


Files produced by the author(s)


  • HAL Id : tel-02487473, version 1


Matthieu Martin. Reconstruction 3D de données échographiques du cerveau du prématuré et segmentation des ventricules cérébraux et thalami par apprentissage supervisé. Imagerie médicale. Université Claude Bernard Lyon 1, 2019. Français. ⟨tel-02487473⟩



Record views


Files downloads