Apparatus and method for denoising medical images

Abstract

Described herein are systems, methods, and apparatuses associated with using deep learning techniques to denoise medical images, such as fluoroscopy images. A first artificial neural network (ANN) is trained to denoise input medical images according to a provided target noise level. The training of the first ANN is performed by pairing noisy input images with target denoised images that include different noise levels. During different training iterations, these target denoised images are generated using a second ANN as an intermediate output of the second ANN. As such, the first ANN can learn to perform the denoising task in an unsupervised manner without requiring noise-free training images as a gold standard.

Description

Technical Field

[0001] This application relates to the field of medical imaging, and more specifically to apparatus and methods for denoising medical images. Background Technology

[0002] Medical images are typically noisy due to constraints imposed on the imaging process and / or the target object, such as acquisition time, radiation dose, and spatial resolution. Various techniques have been developed for denoising medical images. However, because different medical images may contain varying levels of noise (e.g., varying signal-to-noise ratio (SNR)), it is difficult to develop a universal denoising method. Furthermore, even though machine learning (ML) has shown great promise in image processing, the difficulty in obtaining noise-free medical images makes supervised machine learning techniques that require noise-free medical images as the gold standard impractical for medical image denoising. Therefore, there is a strong desire to develop machine learning-based systems and methods that can not only handle medical images with varying noise levels but also do not require supervised training or noise-free medical images as the gold standard. Summary of the Invention

[0003] This document describes systems, methods, and apparatuses associated with denoising medical images, such as fluoroscopic images. An apparatus configured to perform the denoising task described herein may include one or more processors configured to: receive an input medical image including noise, receive a target noise level (e.g., corresponding to a target signal-to-noise ratio (SNR)), and generate an output medical image using a first artificial neural network (ANN) such that at least a portion of the noise is removed from the output medical image according to the target noise level. The first ANN may be trained by a training process to generate the output medical image according to the target noise level. During the training process, the first ANN may be configured to receive a first pair of training images, which may include a first noisy training image and a first target training image, wherein the first target training image may be generated using a second ANN and may include the first noise level. The first ANN may be configured to determine a first noise level associated with the first target training image, generate a denoised version of the first noisy training image using the first ANN based on the first noise level, and adjust the parameters of the first ANN based on the difference (e.g., loss) between the denoised version of the first noisy training image and the first target training image. The first ANN can also be configured to receive a second pair of training images during the training process. The second pair of training images may include a second noisy training image and a second target training image, wherein the second target training image may also be generated using the second ANN and may include a second noise level. The first ANN can determine the second noise level associated with the second target training image, generate a denoised version of the second noisy training image using the first ANN based on the second noise level, and further adjust the parameters of the first ANN based on the difference (e.g., loss) between the denoised version of the second noisy training image and the second target training image.

[0004] In the example, the first noisy training image and the second noisy training image described above may correspond to the same source image, and the first target training image and the second target training image may be generated by the second ANN based on the same source image but with different noise levels (e.g., the first noise level may be different from the second noise level). The second ANN can be pre-trained to denoise the source image through multiple training iterations. During each of the multiple training iterations, the second ANN can be configured to generate a corresponding intermediate denoised output image with a corresponding noise level based on the source image, wherein the first target training image can be obtained as the intermediate denoised output image generated during the first training iteration of the second ANN, and the second target training image can be obtained as the intermediate denoised output image generated during the second training iteration of the second ANN. In the example, the second ANN includes a convolutional neural network (CNN) that is trained to predict pixel values ​​in the output image based on corresponding pixels in the input noisy image and multiple pixels surrounding the corresponding pixels in the input noisy image.

[0005] In the example, the first ANN described above can be configured to process multiple input images through corresponding input channels, and the first ANN can be configured to generate an output medical image by obtaining an image with the same size as the input medical image (wherein, each pixel of the image is assigned a value indicating the target noise level), and processing the obtained image and the input medical image through corresponding first and second input channels of the first ANN. In the example, the first ANN can be configured to shuffle the input medical image into multiple images, and process the multiple images through corresponding input channels of the first ANN, each image having a smaller image size than the input medical image. Attached Figure Description

[0006] The examples disclosed herein can be understood in more detail from the following description, which is given by way of example in conjunction with the accompanying drawings.

[0007] Figure 1 This is a diagram illustrating an example of a neural network system or device that can be used for denoising medical images based on a target noise level.

[0008] Figure 2 This is an example of what can be done with Figure 1 A graph showing example operations associated with the training of a neural network.

[0009] Figure 3 This illustrates the use of a second artificial neural network to generate training data. Figure 1 or Figure 2 An example image of a target denoised image using an artificial neural network.

[0010] Figure 4 This is a diagram illustrating an example of shuffling noisy input images to achieve faster denoising speed.

[0011] Figure 5 This is a flowchart illustrating example operations that can be associated with training a neural network to perform the denoising task described herein.

[0012] Figure 6 This is a block diagram illustrating example components of a device that can be configured to perform one or more tasks described herein. Detailed Implementation

[0013] The present disclosure is illustrated by way of example rather than limitation in the figures.

[0014] Figure 1An example is illustrated of denoising a medical image (e.g., 102) using an artificial neural network (ANN) (e.g., 104) based on a target noise level (e.g., 106). The medical image 102 can be of different types of medical images (e.g., scan images), including, for example, fluorescence fluoroscopy images, magnetic resonance imaging (MRI) images, computed tomography (CT) images, etc. Due to various conditions or constraints associated with the imaging process and / or the objects shown in the image, the medical image 102 may include a certain amount of noise that may affect the quality of the image (e.g., sharpness, clarity, etc.). The ANN 104 can be configured to implement an image denoising model (e.g., a machine learning (ML) model) to remove at least a portion of the noise included in the medical image 102 based on the target noise level 106 and generate an output medical image (e.g., 108) in which the noise level has been reduced to the target noise level 106.

[0015] The target noise level 106 may be pre-configured (e.g., preset) for a denoising system or device configured to implement the denoising model 104. The target noise level 106 may also be received (e.g., in real-time) from a user (e.g., via a user interface) or from an external device (e.g., via a communication interface). In either case, a lower target noise level may cause the denoising model 104 to generate a sharper (e.g., including more high-frequency components) but noisier (e.g., including more artifacts) output medical image, while a higher target noise level may cause the denoising model 104 to generate a blurrier (e.g., including fewer high-frequency components) but cleaner (e.g., including fewer artifacts) output medical image. In the example, a machine learning model may be trained to predict (e.g., based on the noisy image 102) the value of the target noise level 106 that balances sharpness and noise levels in the output image (e.g., predicting the most efficient point between high image sharpness and a low number of noise artifacts). In the example, the target noise level 106 may represent or correspond to the signal-to-noise ratio (SNR).

[0016] ANN 104 can be trained through a training process to learn a denoising model (e.g., this can be achieved through the parameters or weights of ANN 104). Training can be performed using instances of ANNs (e.g., a first ANN), and as will be described in more detail below, ANN training can be performed using images generated by another ANN (e.g., a second ANN), thereby eliminating the need for noise-free images. In the example, either or both of the first and second ANNs can include convolutional neural networks (CNNs), such as fully convolutional neural networks. The CNN can be trained to learn a mapping (e.g., a mathematical mapping) between noisy medical images and clean (e.g., less noisy) medical images based on the information redundancy (e.g., statistical correlation) between pixels in the medical image. In the example, the CNN can be an encoder-decoder network (e.g., with a U-Net structure), which can include multiple convolutional layers and / or one or more pooling layers. Following the convolutional layers may be batch normalization and / or linear or nonlinear activations (e.g., rectified linear units or ReLU activations), and each convolutional layer may include multiple convolutional kernels or filters with corresponding weights, the values ​​of which can be learned through the training process described above to extract features from the input image. The features extracted by the convolutional layers can be downsampled by one or more pooling layers to obtain a representation of the features, e.g., in the form of a feature map or feature vector. The CNN may also include one or more up-pooling layers and one or more transposed convolutional layers. Through the up-pooling layers, the CNN can upsample the features extracted from the input image and further process the upsampled features through one or more transposed convolutional layers (e.g., via multiple deconvolution operations) to derive an amplified (e.g., dense) feature map or feature vector. The amplified feature map or vector can then be used to predict noise-free pixel values ​​(e.g., based on a target noise level of 106) in the output image (e.g., output image 108).

[0017] Figure 2 Examples are shown that can be used with training ANNs (e.g., Figure 1The ANN (104) is used to learn example operations associated with the denoising model described herein. As shown, training can be performed using pairs of training images. Each such pair of training images may include a noisy source training image 202a (e.g., a first noisy training image) and a target denoised image 202b (e.g., a first target training image) with a corresponding noise level. During training (e.g., during training iterations), the noise level (e.g., noise level 206) included in the target denoised image 202b can be determined, for example, by comparing the source training image 202a with the target denoised image 202b. For example, the noise level 206 can be determined by subtracting the target denoised image 202b from the source training image 202a to obtain a noise component, and dividing the power of the target denoised image 202b by the power of the noise component to obtain the noise level of the target denoised image 202b (e.g., as an SNR measurement). The determined noise level (e.g., noise level 206) and the source training image 202a can be fed into ANN 204 to predict a denoised output image (e.g., denoised image 208) with a noise level approximately similar to noise level 206. ANN 204 can then compare the predicted denoised output image (e.g., image 208) with the target denoised image (e.g., image 202b) and determine a loss (e.g., difference) between the two images. The loss (e.g., loss 210) can be determined using various types of loss functions, including, for example, a loss function based on mean squared error (MSE) or a loss function based on structural similarity index (SSIM). Once determined, the loss can be used to tune the parameters of ANN 204 (e.g., the current parameters of the denoising model), for example, by backpropagating the loss through ANN 204 to force the network to adapt its output (e.g., denoised image 208) to the target image (e.g., target denoised image 202b).

[0018] Training of ANN 204 can be repeated in multiple iterations and / or using additional pairs of images. For example, in additional training iterations, ANN 204 can be configured to receive a second pair of training images, which may include a second noisy source image and a second target denoised image. The second noisy source image may be the same as (or different from) the first noisy source image described above (e.g., source image 202a). The second target denoised image may include a noise level different from the first target denoised image (e.g., target denoised image 202b). In response to receiving the second pair of training images, the level of noise included in the second target denoised image can be determined (e.g., using the techniques described above with respect to the first target denoised image). Based on the second source image and the determined noise level, ANN 204 can predict a denoised output image with a noise level approximately similar to the determined noise level. ANN 204 can then compare the predicted denoised output image with the second target denoised image and determine the loss (e.g., difference) between the two images. The loss can be determined using one or more loss functions described herein, and once determined, the loss can be used to further tune the parameters of ANN 204 (e.g., the parameters of the denoising model), for example by backpropagating the loss through ANN 204 to force the network to adapt its output to the second target denoised image.

[0019] Images used to train ANN 204 (such as the first and second target denoised images with corresponding noise levels mentioned above) can be generated using another artificial neural network. Figure 3An example of such an artificial neural network (e.g., ANN 300) is illustrated. As shown, ANN 300 can be a neural network configured to receive a noisy input image I and generate a corresponding output image representing a denoised version of the input image. ANN 300 can be trained to acquire the ability to perform a denoising task over multiple iterations or steps, and during each iteration or step, ANN 300 can generate an output image (e.g., a denoised version of the noisy input image I), which can be an intermediate output image along the way to obtain a final denoised image that satisfies the training objective. Since the difference between the noisy input image and the denoised output image generated during each training iteration or step can decrease with increasing number of training iterations or steps, the intermediate output images generated by ANN 300 from multiple iterations or steps can include corresponding (e.g., different) levels of noise (e.g., corresponding to different denoising intensities). Therefore, these intermediate output images (e.g., different denoised versions of the noisy input image I) can be used as target denoised images during the training of ANN 204, thereby eliminating the need to obtain a noise-free gold standard image. For example, during the various training iterations n1, n2, n3, etc. of the ANN 300, the neural network can generate corresponding intermediate output images In1, In2, In3, etc., which may include noise at corresponding levels (e.g., a specific signal-to-noise ratio (SNR)). These intermediate output images (e.g., In1, In2, In3, etc.) can be paired with the noisy source input image I (e.g., as...). Figure 2 The paired images 202a and 202b shown are used to train the denoising neural network described herein (e.g., Figure 1 ANN 104 or Figure 2 ANN204).

[0020] In the example, ANN 300 may include a convolutional neural network (CNN), such as a fully convolutional neural network, which can be configured to predict pixel values ​​based on one or more surrounding pixels in response to receiving a noisy input image I (e.g., for an output denoised image). These surrounding pixels may come from the same input image (e.g., if a single static image is denoised), or they may come from multiple neighboring images (e.g., if a series of spatial / temporal images, such as images captured in a video, are denoised). The CNN may include multiple convolutional layers and / or one or more pooling layers. Convolutional layers may be followed by batch normalization and / or linear or non-linear activations (e.g., rectified linear units or ReLU activations), and each convolutional layer may include multiple convolutional kernels or filters with corresponding weights, the values ​​of which can be learned through the training process described above to extract features from the input image. The features extracted by the convolutional layers may be downsampled by one or more pooling layers to obtain a representation of the features, e.g., in the form of a feature map or feature vector. The CNN may also include one or more up-pooling layers and one or more transposed convolutional layers. By using an up-pooling layer, a CNN can upsample the features extracted from the input image and further process the upsampled features through one or more transposed convolutional layers (e.g., via multiple deconvolution operations) to derive an amplified (e.g., dense) feature map or feature vector. The amplified feature map or vector can then be used to predict pixel values ​​in an output image that may be free of some noise.

[0021] In the example, deep learning techniques such as Deep Image Prior (DIP) or Generative Adversarial Networks (GANs) can be used to further improve the denoised output image (e.g., to refine high-frequency details in the denoised output image). For example, a dataset of cleaner medical images can be obtained from another imaging modality that is less constrained by imaging time, dose, and / or other factors (e.g., chest X-rays may be cleaner than cardiac X-ray fluoroscopy). Using this cleaner dataset, another ANN can be trained (e.g., with the aforementioned ANNs (such as...). Figure 2 ANN 204 and / or Figure 3 The ANN 204 (or ANN 300) is trained simultaneously to distinguish whether image patches are pruned from the output of the ANN 204 (or ANN 300) or from a cleaner dataset from another imaging modality. Simultaneously, the ANN 204 (or ANN 300) can be trained to make this discrimination task difficult, such that the output image produced by the ANN 204 (or ANN 300) has a patch distribution close to that of the cleaner dataset (e.g., its sharper and cleaner appearance). Using this strategy, an ANN (e.g., ANN 204 or ANN 300) can be trained using cleaner images from one imaging modality (e.g., chest X-ray images) to denoise images produced from another imaging modality (e.g., cardiac X-ray fluoroscopy).

[0022] As described in this article, such as Figure 1 ANN 104 and Figure 2 The ANN 204 denoising neural network can have an encoder-decoder structure (e.g., a U-Net structure), which allows processing of multiple input images through corresponding input channels. Thus, in some examples, in addition to receiving images through the first input channel, such as... Figure 1 In addition to the noisy input image of the noisy image 102 shown, the denoising neural network can also receive representations such as... through the second input channel. Figure 1 The second image, showing a target noise level of 106, is illustrated. This second image can be generated (e.g., via a denoising neural network) to have the same size as the noisy input image, and each pixel in the second image can be assigned a value equal to the target noise level. The two images can then be processed via corresponding first and second input channels to obtain a denoised output image. Thus, when processing individual pixels in the noisy input image, the ANN can access the target noise level through convolution operations. In the example, different pixels in the second image can be assigned different values, which can represent the local target noise level in the noisy input image (e.g., local contrast-to-noise ratio (CNR)). Using this technique, the ANN can preserve details in high-contrast regions and / or eliminate noise in low-contrast regions. The second image can be provided by the user or predicted using another machine learning model.

[0023] In the example, to denoise multiple noisy input images (e.g., which may be part of a video), the denoising neural network described herein (e.g., Figure 1 ANN 104 and Figure 2 The ANN204 can be configured to process multiple input images simultaneously (e.g., using a U-Net structure), allowing temporal information associated with the multiple input images to be used to improve denoising performance. In the example, the denoising neural network described herein may include multiple cascaded subnetworks, each with a U-Net structure to further optimize the denoised output image generated by the neural network. In the example, the noisy input image can be shuffled into multiple smaller noisy images, which can then be processed via corresponding input channels to further improve the denoising effect.

[0024] Figure 4An example is illustrated of shuffling an input noisy image 402 into multiple smaller noisy images 404 to achieve faster processing speed. For example, the input noisy image 402 may have a size of 1024 × 1024 (e.g., in terms of pixels), and the denoising neural network described herein may have the ability to process multiple inputs through four input channels. Thus, the input image 402 can be shuffled into four smaller images (e.g., based on the four input channels) (each smaller image having a size of 512 × 512) before being fed to the denoising neural network described herein, to achieve faster training and inference (e.g., by giving the neural network a sliding window or view of larger image data) and / or better quality. The output generated by the neural network based on the shuffled images can then be shuffled back to its original size and / or shape.

[0025] Figure 5 Examples are shown that can be used with training neural networks (e.g., Figure 1 ANN 104 or Figure 2 The example operations associated with the image denoising task described herein are performed using an ANN (204). As shown, the training operation may include initializing the parameters of the neural network (e.g., weights associated with the individual filters or kernels of the neural network) at 502. The parameters may be initialized, for example, based on samples collected from one or more probability distributions or parameter values ​​from another neural network with a similar architecture. The training operation may also include feeding training data to the neural network at 504 and causing the neural network to predict the denoised image at 506. As described herein, the training data may include pairs of noisy medical images and target denoised images, which may be obtained based on the outputs of individual denoising neural networks at different training steps or training iterations. From the individual target denoised images, the corresponding noise levels can be determined and used at 506 to guide the prediction. At 508, the denoised image predicted by the neural network can be compared with the target denoised image to determine the loss associated with the prediction. The loss can be determined using a suitable loss function, such as a loss function based on MSE, SSIM, etc. Once the loss is determined, it can be evaluated at 510 to determine whether one or more training termination criteria have been met. For example, if the loss is below a predetermined threshold, or if the change in loss between two training iterations (e.g., between consecutive training iterations) falls below a predetermined threshold, then the training termination criterion can be considered satisfied. If it is determined at 510 that the training termination criterion has been satisfied, then training can end. Otherwise, before training returns to 506, the loss can be backpropagated through the neural network at 512 (e.g., based on gradient descent associated with the loss).

[0026] For the sake of simplicity, the training steps are depicted and described in a specific order herein. However, it should be understood that training operations can occur in various orders, simultaneously, and / or with other operations not presented or described herein. Furthermore, it should be noted that not all operations that may be included in the training process are depicted and described herein, and not all exemplified operations need to be performed.

[0027] The systems, methods, and / or apparatuses described herein may be implemented using one or more processors, one or more storage devices, and / or other suitable auxiliary devices (such as display devices, communication devices, input / output devices, etc.). Figure 6 This is a block diagram illustrating an example device 600 that can be configured to perform the image denoising tasks described herein. As shown, device 600 may include a processor (e.g., one or more processors) 602, which may be a central processing unit (CPU), graphics processing unit (GPU), microcontroller, reduced instruction set computer (RISC) processor, application-specific integrated circuit (ASIC), application-specific instruction set processor (ASIP), physical processing unit (PPU), digital signal processor (DSP), field-programmable gate array (FPGA), or any other circuitry or processor capable of performing the functions described herein. Device 600 may also include communication circuitry 604, memory 606, mass storage device 608, input device 610, and / or communication link 612 (e.g., communication bus) through which one or more components shown in the figure exchange information.

[0028] Communication circuitry 604 can be configured to send and receive information using one or more communication protocols (e.g., TCP / IP) and one or more communication networks, including local area networks (LANs), wide area networks (WANs), the Internet, and wireless data networks (e.g., Wi-Fi, 3G, 4G / LTE, or 5G networks). Memory 606 may include a storage medium (e.g., a non-transitory storage medium) configured to store machine-readable instructions that, when executed, cause processor 602 to perform one or more functions described herein. Examples of machine-readable media may include volatile or non-volatile memory, including but not limited to semiconductor memory (e.g., electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM)), flash memory, etc.). Mass storage device 608 may include one or more disks, such as one or more internal hard disks, one or more removable disks, one or more magneto-optical disks, one or more CD-ROMs or DVD-ROMs, etc., on which instructions and / or data may be stored for operation of processor 602. Input device 610 may include a keyboard, mouse, voice-controlled input device, touch-sensitive input device (e.g., touch screen), etc., for receiving user input from device 600.

[0029] It should be noted that device 600 can operate as a standalone device or can be connected to other computing devices (e.g., networked or clustered) to perform the functions described herein. And even in Figure 6 Only one example of each component is shown in the figure, and those skilled in the art will understand that device 600 may include multiple instances of one or more components shown in the figure.

[0030] Although this disclosure has been described according to certain embodiments and generally associated methods, changes and variations of the embodiments and methods will be apparent to those skilled in the art. Therefore, the above description of exemplary embodiments does not limit this disclosure. Other changes, substitutions, and modifications are possible without departing from the spirit and scope of this disclosure. Furthermore, unless specifically stated otherwise, discussions using terms such as “analyze,” “determine,” “enable,” “identify,” and “modify” refer to the actions and processes of a computer system or similar electronic computing device that manipulate and transform data representing physical (e.g., electronic) quantities within the registers and memories of the computer system into other data representing physical quantities within the computer system's memory or other such information storage, transmission, or display devices.

[0031] It should be understood that the above description is intended to be illustrative and not restrictive. Many other embodiments will become apparent to those skilled in the art upon reading and understanding the above description. Therefore, the scope of this disclosure should be determined by reference to the appended claims and the full scope of their equivalents.

Claims

1. A medical image denoising device, comprising: One or more processors, which are configured as follows: Receives input medical images, including those with noise. Received target noise level; And, an output medical image is generated using a first artificial neural network (ANN) such that at least a portion of the noise is removed from the output medical image according to the target noise level, wherein the first ANN is trained through a training process to generate the output medical image according to the target noise level, and wherein, during the training process, the first ANN is configured to: Receive a first pair of training images, which includes a first noisy training image and a first target training image, wherein the first target training image is generated using a second ANN and includes a first noise level; Determine the first noise level associated with the first target training image; The first ANN is used to generate a denoised version of the first noisy training image based on the first noise level; and the parameters of the first ANN are adjusted based on the difference between the denoised version of the first noisy training image and the first target training image. The first ANN is further configured to: adjust the parameters based on the difference between a denoised version of a second noisy training image and a second target training image, the second ANN being pre-trained through multiple training iterations, and in each of the multiple training iterations, the second ANN being configured to generate a corresponding intermediate denoised output image with a corresponding noise level based on the input noisy image, the first target training image being obtained as the intermediate denoised output image generated during a first training iteration of the second ANN, and the second target training image being obtained as the intermediate denoised output image generated during a second training iteration of the second ANN.

2. A method for denoising medical images, the method comprising: Receives input medical images, including those with noise. Received target noise level; And, an output medical image is generated using a first artificial neural network (ANN) such that at least a portion of the noise is removed from the output medical image according to the target noise level, wherein the first ANN is trained through a training process to generate the output medical image according to the target noise level, and wherein, during the training process, the first ANN is configured to: Receive a first pair of training images, which includes a first noisy training image and a first target training image, wherein the first target training image is generated using a second ANN and includes a first noise level; Determine the first noise level associated with the first target training image; The first ANN is used to generate a denoised version of the first noisy training image based on the first noise level; and the parameters of the first ANN are adjusted based on the difference between the denoised version of the first noisy training image and the first target training image. The first ANN is further configured to: adjust the parameters based on the difference between a denoised version of a second noisy training image and a second target training image, the second ANN being pre-trained through multiple training iterations, and in each of the multiple training iterations, the second ANN being configured to generate a corresponding intermediate denoised output image with a corresponding noise level based on the input noisy image, the first target training image being obtained as the intermediate denoised output image generated during a first training iteration of the second ANN, and the second target training image being obtained as the intermediate denoised output image generated during a second training iteration of the second ANN.

3. The method according to claim 2, wherein, During the training process, the first ANN is also configured to: Receive a second pair of training images, which includes a second noisy training image and a second target training image, wherein the second target training image is generated using the second ANN and includes a second noise level; Determine the second noise level associated with the second target training image; Based on the second noise level, the first ANN is used to generate a denoised version of the second noisy training image; and The parameters of the first ANN are adjusted based on the difference between the denoised version of the second noisy training image and the second target training image.

4. The method according to claim 3, wherein, The first noisy training image and the second noisy training image correspond to the same source image, and wherein the first noise level is different from the second noise level.

5. The method according to claim 4, wherein, The first target training image and the second target training image are generated using the second ANN based on the source image.

6. The method according to claim 2, wherein, The first ANN is configured to process multiple input images through corresponding input channels, and wherein using the first ANN to generate the output medical image includes: Obtain an image with the same size as the input medical image, wherein each pixel of the image is assigned a value indicating the target noise level; and The obtained image and the input medical image are processed via the corresponding first and second input channels of the first ANN, or... The first ANN is configured to process multiple input images through corresponding input channels, and generating the output medical image using the first ANN includes: shuffling the input medical images into multiple images, and processing the multiple images through the corresponding input channels of the first ANN, each image having a smaller image size than the input medical image.

7. The method according to claim 2, wherein, The target noise level corresponds to the target signal-to-noise ratio (SNR), and the input medical image includes a fluorescence fluoroscopic image.

8. A computer program product comprising a non-transitory computer-readable storage medium storing a computer program for causing a computer to perform the method of any one of claims 2-7.

9. A method for training a first artificial neural network (ANN) to denoise medical images, the method comprising: A first pair of training images is provided to the first ANN, wherein the first pair of training images includes a noisy input training image and a first target training image, wherein the first target training image is generated by the second ANN from the first training iteration of the second ANN and includes a first noise level; Determine the first noise level associated with the first target training image; The first ANN is used to generate a denoised version of the noisy input training image based on the first noise level; and the parameters of the first ANN are adjusted based on the difference between the denoised version of the noisy input training image and the first target training image. The parameters are adjusted based on the difference between the denoised version of the second noisy training image and the second target training image. The second ANN is pre-trained through multiple training iterations. During each of the multiple training iterations, the second ANN is configured to generate a corresponding intermediate denoised output image with a corresponding noise level based on the input noisy image. The first target training image is obtained as the intermediate denoised output image generated during the first training iteration of the second ANN, and the second target training image is obtained as the intermediate denoised output image generated during the second training iteration of the second ANN.

Images