Volumétrie des ventricules latéraux chez le nouveau-né par segmentation automatique d'échographies 3D

New matrix-array ultrasound (US) probes allow neuroradiologists to acquire volumetric images almost instantly with no sweep of the region of interest. Compared to traditional 2D protocols, 3D US imaging decreases acquisition time without reducing image quality and could increase interpretation capabilities. Monitoring of the brain and lateral ventricles development is necessary especially in cases of premature birth, intraventricular hemorrhage (IVH) and surgical interventions. Significant ventricular dilatation is associated with some neurological disorders as well as lower scores on the Bayley scale of infant development and in some circumstances lower intelligence quotient (IQ). Furthermore, it is also associated with altered white matter development and abnormal volume of cortical gray matter. By monitoring the patients' lateral ventricular dilatation, it is possible to determine if this is a risk factor for their cognitive development or if a surgical intervention is necessary in serious situations. If an abnormality is found with standard 2D US examinations, an MRI can be prescribed for a thorough examination. MRI is challenging with newborns due to immobilization issues, which requires most of the time sedation of the newborn. An alternative is to use recent matrix-array probes instead to perform non-invasive brain imaging through the fontanel. This will allow to perform volumetric analysis with an imaging method more accessible and less expensive than MRI. The project hypothesis is that it is possible to evaluate brain and ventricular development with the 3D US images and accomplish a series of quantitative volumetric measurements. The objective of this project is to validate the volumetric measurement of the 3D US images with the reference MRI and to develop a method to automatically extract the brain volume and segment the lateral ventricles in 3D US. The lateral ventricles volume is important to assess the progression of the dilatation before and after surgical interventions and to assess the severity of the dilatation. First, MRI and 3D US images are acquired for an initial cohort of 12 patients and the lateral ventricles are segmented manually in both imaging modalities. A geometric method is also developed in order to estimate the brain volume which is not fully captured by the limited US probe beam. This method uses an ellipsoid to model the brain shape where its volume is calculated with the 3 ellipsoid semi-axes. This brain volume estimation is compared to the head circumference (HC) which is a widely used method in clinical practice to follow brain development, although there are limitations associated with this approach. Ventricular-brain volume ratio is also calculated to assess the severity of the ventricular dilatation relatively to the brain size. A comparative study and statistical analysis are then undertaken to validate volumes obtained from 3D US images with those from MRI. Results show no statistically significant differences between the extracted MRI and 3D US volumes. Lateral ventricles have a near perfect correlation (r=0.999) and there is an excellent correlation for the brain volume (r=0.988). The difference in volume ratios was 6.0 ± 4.8% compared to MRI. Those analysis are possible on newborns and infants until they are approximately 8 months old, which is the age where the fontanelle starts to close, reducing the acoustic waves propagation. Secondly, the brain and lateral ventricles volumes are automatically extracted from the 3D US images. The brain volume is estimated with the same ellipsoid method after it has been aligned and stripped from the skull. The lateral ventricles were segmented on 13 patients using a multi-atlas registration pipeline with MRI images. Since this is a multimodal registration, a highly specific metric is used to register the MRI with the US images, the LC2 metric (Linear Correlation of Linear Combination). Then, the best registrations are selected for a label fusion but the LC2 alone doesn't allow to automatically select the best registrations between several MRI images. An area weighting term is combined with the LC2 in order to improve the affine registration and to compare the registration results between several MRI images. The area weighting term assigns a weight to each voxel of the projected venricular area based on the position and intensity of the voxel on the US image. Indeed, the ventricular areas are divided in two areas, the fluid cavities which are hypoechoic and the plexus choroïd which is hyperechoic. These regions are used in the calculation of the weighting term. Two algorithms are tested for the label fusion, Majority Voting (MV) and STAPLE. Furthermore, the mesh is refined using deformable mesh model with an energy minimization process. The segmentation results are encouraging (DICE: 70.8±3.6, Mean Absolute Distance: 0.88± 0.20) and the extracted volumes have no statistically significant differences with the manual segmentations. The brain volumes have a mean absolute error with MRI volumes of 7.73% and a good correlation (r=0.942) when automatically segmented. As a comparison, the error was of 3.12% and the correlation excellent (r=0.988) with the manual measurements. In addition, the automatically extracted lateral ventricles volumes have a good correlation (r=0.848) with the manual segmentations and a mean absolute error of 9.84%. The methodology and results presented in this thesis show new perspectives and tools to help evaluate the infants' brain development. This project demonstrates the potential of using new matrix-array US probes to assess brain and lateral ventricular volumes in newborns and infants which could be useful to facilitate monitoring of the lateral ventricles dilatation used for the macrocephaly diagnosis. Furthermore, it is possible to calculate the ventricular-brain volume ratio to assess the dilatation severity relatively to the brain volume.