high-reSolutiON cerebral blood flow estimATIoN basEd on ultrafast ultrasound imaging – SONATINE

Brain surgery is the usual treatment for most brain tumors. After the craniotomy procedure to access the brain, the high-precision removal of the tumor necessitates an accurate definition of the boundary between the tumor and other vital brain tissues. This assessment is usually done visually by the surgeon with the help of medical imaging techniques, such as ultrafast ultrasound imaging (UUI), which has recently become more widespread in clinical practice. However, the precise demarcation of the boundary remains a complex task mainly because of the high vascularization of the peritumoral area, characterized by the proliferation of blood vessels including tiny vessels. Accurately determining the microvasculature of the peritumoral area through UUI-based high-sensitivity and high-resolution blood flow estimation is thus gaining interest. The key idea of such an estimation is to remove unwanted clutter signals from tissues to reveal blood flow as clearly as possible. The robust principal component analysis (RPCA)-based techniques are the most widespread approaches that consist in formulating a mathematical model of the data and then finding a solution by optimizing the related problem. This allows the use of regularizations, e.g., blood sparsity or tissue low rank, to leverage their prior knowledge for efficient blood flow reconstruction. To further improve the reconstruction, a deconvolution was also embedded in the RPCA framework. Despite their efficiency, such techniques still suffer from deadlocks within the above mathematical model and those with associated hyperparameters. This project is dedicated to breaking these deadlocks by taking into account the presence of tissue motion, deriving better non-convex exact relaxations of the relevant optimization problem, and designing a Bayesian-based statistical algorithm for hyperparameter estimation. Then, the resulting techniques will be used to derive a based on a model- or physics-driven approach not only to handle the hyperparameters’ issue but also to yield significantly accurate blood flow estimation, with an additional advantage of an interpretable machine-learning model.