Ultrasound speckle reduction and image reconstruction using deep learning techniques
Ultrasound B-mode images are reconstructed directly from transducer channel signals using a convolutional neural network (CNN). The CNN is trained with a dataset including, as inputs, simulated transducer array channel signals containing simulated speckle and, as outputs, corresponding simulated speckle-free B-mode ground truth images. After training, measured real-time RF signals taken directly from an ultrasound transducer array elements prior to summation are input to the CNN and processed by the CNN to generate as output an estimated real-time B-mode image with reduced speckle.
1. A method for ultrasound image reconstruction using a convolutional neural network (CNN), the method comprising:
(a) training the CNN with a dataset comprising simulated transducer array channel signals containing simulated speckle as inputs, and corresponding simulated speckle-free echogenicity estimates (B-mode images) as outputs; wherein the training uses a loss function involving a norm between estimated B-mode images and simulated speckle-free B-mode images; wherein the estimated B-mode images are estimated by the CNN from the simulated transducer array channel signals;
(b) measuring real-time RF signals taken directly from ultrasound transducer array elements prior to summation;
(c) preprocessing the measured real-time RF signals to apply time delays to focus the array at field points, and inputting the pre-processed measured real-time RF signals to the CNN;
(d) processing by the CNN the pre-processed measured real-time RF signals to produce as output an estimated real-time B-mode image with reduced speckle.
2. The method of claim 1 wherein the loss function is a log-domain normalization-independent loss function.
3. The method of claim 1 wherein the loss function is a normalization-independent mixture of l 1 and multi-scale structural similarity losses.
4. The method of claim 1 wherein the CNN has 8 to 16 convolution blocks, wherein each of the convolution blocks has a 2D convolution, batch normalization, and a rectified linear unit.
5. The method of claim 1 wherein the CNN has 16 layers of convolution blocks and 32 filters per layer.
6. The method of claim 1 further comprising performing a conventional DAS envelope-detection reconstruction to generate a conventional B-mode image from the real-time RF signals, and concatenating the generated conventional B-mode image to an output of a block of the CNN.
7. The method of claim 1 wherein simulated transducer array channel signals are unfocused or focused channel signals.
8. The method of claim 1 wherein the input signals are modulated RF signals or demodulated baseband signals.
9. The method of claim 1 wherein the CNN concatenates the original input data to an intermediate hidden layer of the network.
10. The method of claim 1 wherein the CNN performs a multi-scale or multi-resolution processing of the data via striding or pooling.
11. The method of claim 1 wherein the CNN utilizes 3D convolutions.