Task-oriented deep learning image denoising
In one embodiment, there is provided an apparatus for denoising a medical image. The apparatus includes a denoising artificial neural network (ANN) configured to denoise input image data. The denoising ANN is trained, based at least in part, on at least one loss function. The at least one loss function includes a task-oriented loss.
1 . An apparatus for denoising a medical image, the apparatus comprising:
a computer readable storage device;
a denoising artificial neural network (ANN) configured to denoise input image data;
a training module, configured to train the denoising ANN according to instructions stored on the computer-readable storage device, based at least in part, on at least one loss function, the at least one loss function comprising a task-oriented loss function,
wherein the task-oriented loss function is related to a downstream task,
wherein the training module comprises a task-representative network corresponding to the task-oriented loss function,
wherein the task-representative network is configured to be pretrained according to instructions stored on the computer readable storage device with task pretraining data, the task pretraining data including pretraining input data and corresponding pretraining target output data, and
wherein pre-training the task-representative network includes determining and fixing task-representative network parameters related to the downstream task so that the task network parameters are not adjusted during training the denoising ANN.
2 . The apparatus of claim 1 , wherein the denoising ANN corresponds to a generator of a Wasserstein generative adversarial network (WGAN).
3 . The apparatus of claim 1 , wherein the task-oriented loss is related to image segmentation.
4 . The apparatus of claim 1 , wherein the denoising ANN is trained based, at least in part, on a plurality of loss functions, the plurality of loss functions comprising the task-oriented loss function and at least one of a mean square error (MSE), and a discriminator loss corresponding to a Wasserstein generative adversarial network (WGAN).
5 . The apparatus of claim 1 , wherein the input image data is selected from the group comprising computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET), and single-photon emission computerized tomography (SPECT).
6 . The apparatus of claim 1 , wherein the task-oriented loss corresponds to a training task-representative network.
7 . The apparatus of claim 6 , wherein an architecture of the training task-representative ANN is different from an architecture of an actual task-representative ANN.
8 . A method for denoising a medical image, the method comprising:
pretraining a training module comprising a task-representative network corresponding to a task-oriented loss function with task pretraining data, the task pretraining data including pretraining input data and corresponding pretraining target output data, wherein pre-training the task-representative network includes determining and fixing task-representative network parameters related to the downstream task so that the task network parameters are not adjusted during training the denoising ANN;
training, by the training module, a denoising artificial neural network (ANN), based at least in part, on at least one loss function, the at least one loss function comprising a task-oriented loss function; and
denoising, by the denoising ANN, input image data,
wherein the task-oriented loss function is related to a downstream application.
9 . The method of claim 8 , wherein the denoising ANN corresponds to a generator of a Wasserstein generative adversarial network (WGAN).
10 . The method of claim 8 , wherein the task-oriented loss is related to image segmentation.
11 . The method of claim 8 , wherein the input image data is selected from the group comprising computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET), and single-photon emission computerized tomography (SPECT).
12 . The method of claim 8 , wherein the training module further comprises at least one of a mean square error (MSE) loss function, and a discriminator corresponding to a Wasserstein generative adversarial network (WGAN), and the denoising ANN is trained based, at least in part, on a plurality of loss functions.
13 . A computer readable storage device having stored thereon instructions that when executed by one or more processors result in the operations comprising the method according to claim 8 .
14 . A system for denoising a medical image, the system comprising:
a computing device comprising a processor, a memory, an input/output circuitry, and a data store;
a denoising artificial neural network (ANN) configured to denoise input image data;
a training module, configured to train the denoising ANN based at least in part on at least one loss function, the at least one loss function comprising a task-oriented loss function,
wherein the task-oriented loss function is related to a downstream application, and
wherein the training module comprises a task-representative network corresponding to the task-oriented loss function, the training module configured to be pre-trained with task pretraining data, the task pretraining data including pretraining input data and corresponding pretraining target output data, and
wherein pre-training the task-representative network includes determining and fixing task-representative network parameters related to the downstream task so that the task network parameters are not adjusted during training the denoising ANN.
15 . The system of claim 14 , wherein the denoising ANN corresponds to a generator of a Wasserstein generative adversarial network (WGAN).
16 . The system of claim 14 , wherein the task-oriented loss is related to image segmentation.
17 . The system of claim 14 , wherein the input image data is selected from the group comprising computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET), and single-photon emission computerized tomography (SPECT).
18 . The system of claim 14 , wherein the training module further comprises at least one of a mean square error (MSE) loss function, and a discriminator corresponding to a Wasserstein generative adversarial network (WGAN), and the denoising ANN is trained based, at least in part, on a plurality of loss functions.