3-D convolutional autoencoder for low-dose CT via transfer learning from a 2-D trained network
A 3-D convolutional autoencoder for low-dose CT via transfer learning from a 2-D trained network is described, A machine learning method for low dose computed tomography (LDCT) image correction is provided. The method includes training, by a training circuitry, a neural network (NN) based, at least in part, on two-dimensional (2-D) training data. The 2-D training data includes a plurality of 2-D training image pairs. Each 2-D image pair includes one training input image and one corresponding target output image. The training includes adjusting at least one of a plurality of 2-D weights based, at least in part, on an objective function. The method further includes refining, by the training circuitry, the NN based, at least in part, on three-dimensional (3-D) training data. The 3-D training data includes a plurality of 3-D training image pairs. Each 3-D training image pair includes a plurality of adjacent 2-D training input images and at least one corresponding target output image. The refining includes adjusting at least one of a plurality of 3-D weights based, at least in part, on the plurality of 2-D weights and based, at least in part, on the objective function. The plurality of 2-D weights includes the at least one adjusted 2-D weight.
1. A method for low dose computed tomography (LDCT) image correction, the method comprising:
training, by a training circuitry, a convolutional neural network (CNN) based, at least in part, on two-dimensional (2-D) training data, the 2-D training data comprising a plurality of 2-D training image pairs, each 2-D image pair comprising one training input image and one corresponding training output image, the training comprising adjusting at least one of a plurality of 2-D filter parameters based, at least in part, on an objective function; and;
refining, by the training circuitry, the CNN based, at least in part, on three-dimensional (3-D) training data, the 3-D training data comprising a plurality of 3-D training image pairs, each 3-D training image pair comprising a plurality of adjacent 2-D training input images and at least one corresponding training output image, the refining comprising adjusting at least one of a plurality of 3-D filter parameters based, at least in part, on the plurality of 2-D filter parameters and based, at least in part, on the objective function, the plurality of 2-D filter parameters comprising the at least one adjusted 2-D filter parameter;
wherein the CNN is a contracting path-based convolutional auto encoder (CPCA) comprising a plurality of main convolutional layers, a plurality of deconvolutional layers and a plurality of contracting paths, each contracting path coupled between an output of a respective convolutional layer and a respective selected deconvolutional layer, each of the plurality of contracting paths comprises a respective contracting convolutional layer configured to receive as input a respective first feature map from a respective selected main convolutional layer and to output a respective second feature map to a respective selected deconvolutional layer, the respective second feature map being different than the respective first feature map.
2. The method of claim 1 , wherein the CNN corresponds to a generator network of a Wasserstein generative adversarial network (WGAN) with gradient penalty.
3. The method of claim 1 , wherein the objective function comprises an adversarial loss function and a perceptual loss function.
4. The method of claim 1 , wherein each 3-D training image pair comprises three adjacent 2-D training input images.
5. The method of claim 1 , wherein the image correction corresponds to denoising an LDCT image to approximate a corresponding normal dose CT (NDCT) image.
6. The method of claim 1 , further comprising correcting, by the trained CNN, an actual LDCT image to approximate a corresponding actual normal dose CT (NDCT) image, the correcting comprising at least one of noise reduction and/or artifact reduction.
7. A convolutional neural network (CNN) stored on a non-transitory computer readable storage medium and configured to be executed by one or more processors of the non-transitory computer readable storage medium, the CNN comprising:
a contracting path-based convolutional auto encoder (CPCA) comprising:
a plurality of main convolutional layers coupled in series;
a plurality of deconvolutional layers coupled in series; and
a plurality of contracting paths, each contracting path coupled between an output of a respective convolutional layer and a respective selected deconvolutional layer, each of the plurality of contracting paths comprises a respective contracting convolutional layer configured to receive as input a respective first feature map from a respective selected main convolutional layer and to output a respective second feature map to a respective selected deconvolutional layer, the respective second feature map being different than the respective first feature map;
wherein the CPCA is trained based, at least in part, on two-dimensional (2-D) training data, the 2-D training data comprising a plurality of 2-D training image pairs, each 2-D image pair comprising one training input image and one corresponding training output image, the training comprising adjusting at least one of a plurality of 2-D filter parameters based, at least in part, on an objective function and the CPCA is refined based, at least in part, on three-dimensional (3-D) training data, the 3-D training data comprising a plurality of 3-D training image pairs, each 3-D training image pair comprising a plurality of adjacent 2-D training input images and at least one corresponding training output image, the refining comprising adjusting at least one of a plurality of 3-D filter parameters based, at least in part, on the plurality of 2-D filter parameters and based, at least in part, on the objective function, the plurality of 2-D filter parameters comprising the at least one adjusted 2-D filter parameter.
8. The CNN of claim 7 , wherein the CPCA corresponds to a generator network of a Wasserstein generative adversarial network (WGAN) with gradient penalty.
9. The CNN of claim 7 , wherein the objective function comprises an adversarial loss function and a perceptual loss function.
10. The CNN of claim 7 , wherein the CPCA is configured to correct an actual low dose computed tomography (LDCT) image to approximate a corresponding actual normal dose computed tomography (NDCT) image, the correcting comprising at least one of noise reduction and/or artifact reduction.
11. The CNN according to claim 7 , wherein the CPCA comprises a first main convolutional layer, a second main convolutional layer, a third main convolutional layer and a fourth main convolutional layer; a first deconvolutional layer, a second deconvolutional layer, a third deconvolutional layer and a fourth deconvolutional layer; and a first contracting path coupling an output of the first main convolutional layer to the fourth deconvolutional layer, a second contracting path coupling an output of the second main convolutional layer to the third deconvolutional layer, and a third contracting path coupling an output of the third main convolutional layer to the second deconvolutional layer.
12. The CNN according to claim 7 , further comprising a plurality of rectified linear units (ReLUs), each ReLU coupled an output of a respective main convolutional layer or an output of a respective deconvolutional layer.
13. A low dose computed tomography (LDCT) image correction system comprising at least one device arranged to perform the method of claim 1 .
14. A low dose computed tomography (LDCT) image correction device comprising means to perform the method of claim 1 .
15. A computer readable storage device having stored thereon instructions that when executed by one or more processors result in the following operations comprising the method according to claim 1 .