Artifact Suppression for Passive Cavitation Imaging Using U-Net CNNs with Uncertainty Quantification

Abstract

The clinical applications of chemotherapy treatments for brain tumors are limited by the blood-brain barrier (BBB) which blocks drugs from reaching the tumors. Magnetic resonance imaging (MRI)-guided focused ultrasound (FUS) is applied to the brain to create micro-bubbles in the BBB for opening the BBB temporarily for targeted drug delivery. Creation of these microbubbles is monitored using passive cavitation imaging (PCI). Generally, PCI utilizes delay-and-sum (DAS) beam-forming, which suffers from a space-varying point spread function and high sidelobes both of which introduce image artifacts. To address this issue and to better monitor the BBB opening, a deep learning denoising approach is proposed based on a U-Net convolutional neural networks (CNN) architecture. The U-Net is trained with pairs of images, with DAS passive cavitation images at the input layer, and ground true noise-free images at the output layer. In the training phase, the network fits the mapping from the DAS images to the noise-free images. A Monte Carlo (MC) dropout approach is used to quantify uncertainty in de-noised images, giving information that can be further used to suppress artifacts. It is shown in this work that the resulting model can generate de-noised DAS images, and that uncertainty quantification in U-net output can be used to further improve artifact suppression.