Functional ultrasound imaging of the brain using deep learning and sparse data
View Patent ↗A method for ultrasound power Doppler image reconstruction includes obtaining sparse sequences of compound frames of ultrasound data; inputting the sparse sequences into a convolutional neural network; and generating as output from the convolutional neural network a reconstructed time series of power Doppler images corresponding to the sparse sequences. The convolutional neural network is trained using a custom loss function to learn a reconstruction function using ground truth images from high-quality in vivo images.
1. A method for ultrasound power Doppler image reconstruction comprising:
obtaining sparse temporal sequences of compound frames of ultrasound data, wherein the sparse temporal sequences are undersampled by at least 75% compared with fully-sampled data;
inputting the sparse temporal sequences into a convolutional neural network;
generating as output from the convolutional neural network a reconstructed time series of power Doppler images corresponding to the sparse temporal sequences;
wherein the convolutional neural network is trained to learn a power Doppler reconstruction function that maps a sequence of compound frames to a corresponding power Doppler image, wherein the convolutional neural network generates the power Doppler image as output at a final layer of the convolutional neural network from the sequence of compound frames input at a first layer of the convolutional neural network, wherein the convolutional neural network is trained using ground truth images from high-quality in vivo images, wherein the convolutional neural network is trained using a custom loss function.
2. The method of claim 1 wherein the convolutional neural network is a U-Net with drop-out layers.
3. The method of claim 1 wherein the convolutional neural network is a U-Net with an input layer of 3D convolutional filters that extract spatiotemporal features from the sparse temporal sequences.
4. The method of claim 1 wherein the custom loss function is defined as a weighted sum of 1) a mean absolute error between a predicted Deep-fUS image and a respective ground truth (LMAE) and 2) a structural dissimilarity index metric loss (LSSIM).
5. The method of claim 1 wherein the ultrasound power Doppler image reconstruction is performed in real-time intrasurgically for guidance and monitoring.
6. The method of claim 1 wherein the ultrasound power Doppler image reconstruction is performed in real-time in procedures in newborn through the fontanel window by reducing data acquisition, storage, and processing resources, and by reducing sensitivity to motion artifacts.