WO2019074938A1 - Contrast dose reduction for medical imaging using deep learning - Google Patents

Contrast dose reduction for medical imaging using deep learning Download PDF

Info

Publication number
WO2019074938A1
WO2019074938A1 PCT/US2018/055034 US2018055034W WO2019074938A1 WO 2019074938 A1 WO2019074938 A1 WO 2019074938A1 US 2018055034 W US2018055034 W US 2018055034W WO 2019074938 A1 WO2019074938 A1 WO 2019074938A1
Authority
WO
WIPO (PCT)
Prior art keywords
contrast
images
image
dose
full
Prior art date
Application number
PCT/US2018/055034
Other languages
French (fr)
Inventor
Greg Zaharchuk
Enhao GONG
John M. Pauly
Original Assignee
The Board Of Trustees Of The Leland Stanford Junior University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by The Board Of Trustees Of The Leland Stanford Junior University filed Critical The Board Of Trustees Of The Leland Stanford Junior University
Priority to AU2018346938A priority Critical patent/AU2018346938B2/en
Priority to CA3078728A priority patent/CA3078728A1/en
Priority to CN201880072487.0A priority patent/CN111601550B/en
Priority to JP2020520027A priority patent/JP7244499B2/en
Priority to KR1020207013155A priority patent/KR20200063222A/en
Priority to BR112020007105-6A priority patent/BR112020007105A2/en
Priority to EP18866524.4A priority patent/EP3694413A4/en
Priority to SG11202003232VA priority patent/SG11202003232VA/en
Publication of WO2019074938A1 publication Critical patent/WO2019074938A1/en
Priority to JP2023036252A priority patent/JP7476382B2/en

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/60—Rotation of whole images or parts thereof
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/60—Image enhancement or restoration using machine learning, e.g. neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/50—Depth or shape recovery
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10088—Magnetic resonance imaging [MRI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10116—X-ray image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10116—X-ray image
    • G06T2207/10121—Fluoroscopy
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10132—Ultrasound image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20212—Image combination
    • G06T2207/20221—Image fusion; Image merging
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30016—Brain
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems

Definitions

  • This invention relates generally to medical diagnostic imaging. More specifically, it relates to imaging techniques that use contrast agents.
  • contrast agents to enhance the visualization of normal and abnormal structures. Examples include conventional angiography, fluoroscopy, computed tomography (CT), ultrasound, and magnetic resonance imaging (MRI). It is often desirable to reduce the contrast agent dose in order to increase the safety of these agents. However, reduced dose reduces desired imaging enhancements, so this has not been possible.
  • MRI is a powerful imaging technique providing unique information to distinguish different soft tissues and pathologies. Magnetic contrast agents with unique relaxation parameters are often administered to further boost the visibility of pathology and delineation of lesions.
  • Gadolinium based contrast agents are widely used in MRI exams because of their paramagnetic properties, for applications such as angiography, neuroimaging and liver imaging.
  • CE-MRI contrast enhanced MRI
  • the present invention provides techniques to enhance image quality of diagnostic imaging modalities using a lower dose of contrast than is currently possible. This enables new opportunities for improving the value of medical imaging.
  • the techniques are generally applicable to a variety of diagnostic imaging techniques including angiography, fluoroscopy, computed tomography (CT), and ultrasound.
  • the techniques of the present invention are able to predict a synthesized full-dose contrast agent image from a low-dose contrast agent image and a pre-dose image.
  • the low dose may be any fraction of the full dose, but is preferably 1/10 or less of the full dose.
  • Significantly, naively amplifying the contrast enhancement of a 1/10 low-dose CE-MRI by a factor of ten results in poor image quality with widespread noise and ambiguous structures.
  • the techniques of the present invention remarkably are able to recover the full contrast signal and generate predicted full-contrast images with high diagnostic quality.
  • Embodiments of the invention use a deep learning network, such as a Convolutional Neural Network for image-to-image regression, with a pre-contrast image and low-contrast image as input, and with a predicted full-contrast image as output.
  • a residual learning approach is preferably used in prediction.
  • the method includes preprocessing to co-register and normalize between different images so they are directly comparable. This step is important since there are arbitrary different acquisition and scaling factor for each scan. Preferably, the average signal is used for normalization.
  • the preprocessed images are then used to train the deep learning network to predict the full-contrast image from the pre-contrast and low-contrast images.
  • the trained network is then used to synthesize full-contrast images from clinical scans of pre-contrast and low-contrast images.
  • the techniques are generally applicable for any diagnostic imaging modality that uses a contrast agent. For example, imaging applications would include fluoroscopy, MRI, CT, and ultrasound.
  • the invention provides a method for training a diagnostic imaging device to perform medical diagnostic imaging with reduced contrast agent dose.
  • the method includes a) performing diagnostic imaging of a set of subjects to produce a set of images comprising, for each subject of the set of subjects, i) a full-contrast image acquired with a full contrast agent dose administered to the subject, ii) a low-contrast image acquired with a low contrast agent dose administered to the subject, where the low contrast agent dose is less than the full contrast agent dose, and iii) a zero-contrast image acquired with no contrast agent dose administered to the subject; b) pre-processing the set of images to co-register and normalize the set of images to adjust for acquisition and scaling differences between different scans; and c) training a deep learning network (DLN) with the pre-processed set of images by applying zero-contrast images from the set of images and low-contrast images from the set of images as input to the DLN and using a cost function to compare the output of the DLN
  • the invention provides a method for medical diagnostic imaging with reduced contrast agent dose.
  • the method includes a) performing diagnostic imaging of a subject to produce a low-contrast image acquired with a low contrast agent dose administered to the subject, where the low contrast agent dose is less than a full contrast agent dose, and a zero- contrast image acquired with no contrast agent dose administered to the subject; b) preprocessing the low-contrast image and zero-contrast image to co-register and normalize the images to adjust for acquisition and scaling differences; and c) applying the low-contrast image and the zero-contrast image as input to a deep learning network (DLN) to generate as output of the DLN a synthesized full-dose contrast agent image of the subject; where the DLN has been trained by applying zero-contrast images and low-contrast images as input and full-contrast images as reference ground-truth images.
  • DLN deep learning network
  • the low contrast agent dose is preferably less than 10% of a full contrast agent dose.
  • the diagnostic imaging may be angiography, fluoroscopy, computed tomography (CT), ultrasound, or magnetic resonance imaging.
  • performing diagnostic imaging may include performing magnetic resonance imaging where the full contrast agent dose is at most 0.1 mmol/kg Gadolinium MRI contrast.
  • the DLN is an encoder-decoder convolutional neural network (CNN) including bypass concatenate connections and residual connections.
  • CNN encoder-decoder convolutional neural network
  • the invention provides a method for medical diagnostic imaging including a) performing diagnostic imaging of a subject to produce a first image acquired with a first image acquisition sequence and a second image acquired with a second image acquisition sequence distinct from the first image acquisition sequence, where zero contrast agent dose is administered during the diagnostic imaging; b) pre-processing the first image and the second image to co-register and normalize the images to adjust for acquisition and scaling differences; c) applying the first image and the second image as input to a deep learning network (DLN) to generate as output of the DLN a synthesized full-dose contrast agent image of the subject; wherein the DLN has been trained by applying zero-contrast images with different imaging sequences as input and full-contrast images acquired with a full contrast agent dose as reference ground-truth images.
  • DLN deep learning network
  • FIG. 1 is a flow chart showing a processing pipeline according to an embodiment of the invention.
  • FIG. 2 illustrates a workflow of a protocol and procedure for acquisition of images used for training according to an embodiment of the invention.
  • FIG. 3 is a schematic overview of an image processing flow according to an embodiment of the invention.
  • FIG. 4 is an illustration of a signal model to synthesize full-dose contrast-enhanced MRI image according to an embodiment of the invention.
  • FIG. 5 shows a detailed deep learning (DL) model architecture according to an embodiment of the invention.
  • FIG. 6 is a set of images of a patient with intracranial metastatic disease, showing the predicted image synthesized from pre-contrast image and low-dose image, compared with a full-dose image according to an embodiment of the invention.
  • DL deep learning
  • FIG. 7 is a set of images of a patient with a brain neoplasm, showing a synthesized full-dose image predicted from a pre-contrast image and low-dose image, compared with a full-dose image according to an embodiment of the invention.
  • FIG. 8 is a set of images of a patient with a programmable shunt and intracranial hypotension, showing noise suppression capability and consistent contrast enhancement capability according to an embodiment of the invention.
  • Embodiments of the present invention provide a deep learning based diagnostic imaging technique to significantly reduce contrast agent dose levels while maintaining diagnostic quality for clinical images.
  • FIG. 1 and FIG. 2 Detailed illustrations of the protocol and procedure of an embodiment of the invention are shown in FIG. 1 and FIG. 2.
  • FIG. 1 and FIG. 2 Detailed illustrations of the protocol and procedure of an embodiment of the invention are shown in FIG. 1 and FIG. 2.
  • the embodiments described below focus on MRI imaging for purposes of illustration, the principles and techniques of the invention described herein are not limited to MRI but are generally applicable to various imaging modalities that make use of contrast agents.
  • FIG. 1 is a flow chart showing a processing pipeline for an embodiment of the invention.
  • a deep learning network is trained using multi -contrast images 100, 102, 104 acquired from scans of a multitude of subjects with a wide range of clinical indications.
  • the images are pre- processed to perform image co-registration 106, to produce multi -contrast images 108, and data augmentation 110 to produce normalized multi -contrast image patches 112.
  • Rigid or non-rigid co-registration may be used to adjust multiple image slices or volumes to match the pixels and voxels to each other. Since there can be arbitrary scaling differences between different volumes, normalization is used to match the intensity of each image/volume. Brain and anatomy masks are used optionally to extract the important regions of interests in each image/volume.
  • Reference images 104 are also processed to perform co-registration and normalization 118. These pre-processed images are then used to train a deep learning network 114, which is preferably implemented using residual learning in a convolutional neural network.
  • the input to the deep learning network is a zero-contrast dose image 100 and low-contrast dose image 102, while the output of the network is a synthesized prediction of a full-contrast dose image 116.
  • a reference full contrast image 104 is compared with the synthesized image 116 using a loss function to train the network using error backpropagation.
  • FIG. 2 illustrates the workflow of the protocol and procedure for acquisition of images used for training.
  • a pre-contrast (zero-dose) image 200 is acquired, a low dose (e.g., 10%) of contrast is administered and a low-dose image 202 is acquired.
  • An additional dose (e.g., 90%) of contrast is then administered to total a full 100% dose, and a full-dose image 204 is then acquired.
  • images are acquired with 3T MRI scanners (GE Healthcare, Waukesha, WI, USA) using standard neuro clinical protocol with high-resolution 3D Tl -weighted inversion-recovery prepped fast-spoiled-gradient-echo (IR-FSPGR) imaging at 3T.
  • IR-FSPGR fast-spoiled-gradient-echo
  • high-resolution Tl -weighted IR-FSPGR pre-contrast images, post-contrast images with 10% low-dose and 100% full-dose of gadobenate dimeglumine (0.01 and 0.1 mmol/kg, respectively) full-dose images are acquired.
  • similar setups are used to include at least one set of images without enhancement that is acquired without injecting contrasts. And optionally at least one set of images with low-level enhancement that is acquired with injecting low dosage of contrasts.
  • the contrast is usually an iodine-based contrast agent.
  • the CT contrast agent is usually administered based on preference of physicians, regulatory standard and patient weight.
  • the administration of a low dose means injecting less than standard protocols.
  • CT there can be multiple set of images with different contrast visualization (possibly the same dosage but different visual appearance) by using multiple energy radiation.
  • the ground-truth is a set of images acquired with 100% full-dose of CT contrast.
  • the contrast agent can be, for example, microbubbles.
  • FIG. 3 is a schematic overview of the image processing flow.
  • the steps of the detailed workflow are applied to the multi -contrast images acquired for each subject.
  • acquisition stage 300 scans are performed as described above to acquire a zero-contrast image 306, low-contrast image 308, and full-contrast image 310.
  • These multi -contrast images then pass through a preprocessing stage 302.
  • the resulting images then are used in a deep learning training stage 304 for training a deep learning network to synthesize a full-dose image 312.
  • Pre-processing steps include image co-registration and signal normalization.
  • Normalization is used to remove bias in the images obtained at different dosage levels, which is performed using a mask 314 of non- enhancing tissues. Specifically, to remove the systematic differences between different signal intensity levels in non-enhancing regions (such as scalp fat), co-registration and signal normalization based on average voxel value within a mask is performed. Alternatively, the normalization can be based on max/min intensity of the images or certain percentiles. Also, the normalization can be based on matching the distribution of intensity to certain predefined or standard distribution. Mask can be applied to calculate the average, max, min, percentile or the distribution to get better normalization performance to match different set of images. This step is performed because the transmit and receive gains used for 3 sequences of the 3 different scans are not guaranteed to be the same. For CT and ultrasound the scaling and normalization is also applicable as there are possible intensity re-scaling steps in both acquisition and image storage process.
  • FIG. 4 is an illustration of the signal model to synthesize full-dose contrast-enhanced MRI image 412.
  • the low-dose contrast uptake 408 between pre-contrast 400 and low-dose 402 is noisy but does include contrast information.
  • a deep learning network is trained using the true 100% full-dose CE-MRI images as the reference ground-truth.
  • the non-contrast (zero-dose) MRI and the 10% low-dose CE-MRI are provided to the network as inputs, and the output of the network is an approximation of the full-dose CE-MRI.
  • this network implicitly learns the guided denoising of the noisy contrast uptake extracted from the difference signal between low- dose and non-contrast (zero-dose) images, which can be scaled to generate the contrast enhancement of a full-dose image.
  • the detailed deep learning (DL) model architecture in one embodiment of the invention is shown in FIG. 5.
  • This model is an encoder-decoder convolutional neural network with 3 encoder steps 500, 502, 504 and 3 decoder steps 506, 508, 510.
  • 3 convolutional layers connected by 3x3 Conv-BN-ReLU.
  • Encoding steps are connected in sequence by 2x2 max-pooling, and decoder steps are connected in sequence by 2x2 up- sampling.
  • Bypass concatenate connections 512, 514, 516 combine symmetric layers to avoid resolution loss.
  • the residual connections 518, 520, 522, 524, 526, 528, 530 enable the model to synthesize a full-dose image by predicting the enhancement signal 540 from a difference between pre-dose image 542 and low-dose image 544.
  • the cost function 532 compares the predicted full-dose image 534 and the reference ground-truth full-dose image 536, which enables the optimization of the network parameters via error backpropagation 538.
  • the network can have different number of layers, image size in each layers and variable connections between layers.
  • the function in each layer can be different linear or nonlinear functions.
  • the output layer can have a different so-called activation function that maps to certain range of output intensity.
  • the network was trained on around 300 2D slices of the co-registered 3D volumes in each patient, excluding the slices at the base of the brain which had low S R and no valuable information for anatomy or contrast. Standard image rigid transformations are used to further augment the dataset in training to ensure the robustness of the model.
  • stochastic gradient descent (SGD) was used for each subset of training datasets and backpropagation is used to optimize the network parameters with respect to a cost function comparing predicted and true full-dose MR images.
  • the mean-absolute- error (MAE) cost function also known as the LI loss, is used in training.
  • Training takes 200 epochs with SGD and ADAM method for optimization and 10% of the training dataset with random permutations was used for validation to optimize hyper-parameters and pick out the best model among all iterations.
  • SGD and ADAM are also examples of operators to solve the optimization problem in training the network. There are many other options including, for example, RMSprop and Adagrad. Basically, the optimizers enable faster and smoother convergence.
  • the loss functions can include, for example, a function map from the ground-truth image and predicted image pair to a set loss values. This includes pixel-wise or voxel-wise loss that is based on pixel/voxel differences which usually use LI (mean-absolute-error) or L2 (mean- squared-error) loss. Also, the loss can be based on regions with certain size that considering similarity of the structures, e.g., SSIM structural similarities. Or the loss can be computed based on other previously or concurrently trained networks that so-called perceptual loss or adversarial loss respectively. Also the loss can be any arbitrary weighted combination of many loss functions.
  • the trained deep learning network is used to synthesize a full-dose image from zero-dose and low-dose images.
  • the co-registered and normalized non-contrast (zero- dose) and low-dose images are loaded from DICOM files and input to the trained network.
  • With efficient forward-passing, which takes around 0.1 sec per 512-by-512 image the synthesized full-dose image is generated.
  • the process is then conducted for each 2D slice to generate entire 3D volume and stored in a new DICOM folder for further evaluation.
  • the predicted image 606 synthesized from pre-contrast image 600 and low-dose image 602 has similar highlighting of contrast enhancement in the lesions as full-dose image 604.
  • the lesions show improved visibility in the synthesized full-contrast image 606 while they cannot be reliably appreciated in low-dose CE-MRI image 602.
  • the synthesized CE-MRI image 606 shows a similar outline of a metastatic lesion in the right posterior cerebellum compared with the one in true full-contrast CE-MRI image 604.
  • results demonstrate better noise suppression capability and consistent contrast enhancement capability.
  • the engorgement of the dural sinuses and pachymeningeal enhancement are clearly seen on both the synthesized image 806, predicted from pre-contrast image 800 and low-dose image 802, and true full-dose image 804, while there is less noise in non-enhancing regions on the synthesized image.
  • the method generates diagnostic quality contrast from the low-dose acquisition, also demonstrating improved noise suppression and contrast enhancement.
  • the synthesized CE-MRI images are improved significantly (both p ⁇ 0.001) compared with the low-dose image on all quantitative metrics with 11.0% improvements in SSIM and over 5.0 dB gains in PS R.
  • the proposed method is not significantly different (both p > 0.05) for overall image quality and the clarity of the enhanced contrast.
  • the 10% low-contrast images are significantly (p ⁇ 0.001) worse in the clarity of the enhanced contrast, with a decrease of over 3 on a 5-point scale.
  • the synthesized full- dose images show significant (p ⁇ 0.05) improvements over the acquired full-dose CE-MRI in suppressing artifacts in the non-enhancing regions, which was an additional advantage of the DL method.
  • the illustrative embodiment described above uses a 2D CNN. Further performance gains are achievable with 3D CNN models considering the correlated spatial information in adjacent thin slices.
  • 2D CNN process single slice of image
  • 2.5D multiple slices of image
  • 3D entire volume
  • patch based patch from volume
  • MAE loss is used as cost function to train the DL model in the above illustrative embodiment, other loss functions are envisioned, such as mixture loss functions with non-local structural similarities or using Generative Adversarial Network (GAN) as a data driven cost function.
  • GAN Generative Adversarial Network
  • a fixed level of gadolinium contrast is used with 90% reduction from the original clinical usage level.
  • full-dose is used herein to refer to a standard dosage enabling certain defined visual quality.
  • contrasts are treated as drug and they have FDA regulated full-dose that should be used.
  • the full-dose means a standard dosage that is recommended or required by the FDA or clinicians to achieve good diagnostic quality.
  • the technique can make use of a patch based method wherein the input image size can be different than 128x128x1. It can be any xxyxz size that uses a patch of size x*y and considering depth/thickness of z. It is also envisioned that low dose levels higher than 10% may be used. In general, it is beneficial to minimize the dose subject to the constraint that the synthesized image retains diagnostic quality. The resulting minimal low dose may vary depending on contrast agent and imaging modality.
  • the network can be trained to automatically generalized to improve the lower dose to the higher dose i3 ⁇ 4igh, where the higher dose level is a full dose or less, and the lower dose level is less than the high dose, i.e., 0 ⁇ ⁇ dhigh ⁇ 1.
  • the inventors also envision synthesis of high dose images from zero dose images, without the need for any contrast dose images.
  • the zero dose images in this case include multiple MR images acquired using different sequences (e.g., Tlw, T2w, FLAIR, DWI) that show different appearance/intensity of different tissues.
  • the trained network can then predict a synthesized full- contrast Tlw to be used for diagnosis.
  • the deep learning network is trained with a first image acquired with a first image acquisition sequence and a second image acquired with a second image acquisition sequence distinct from the first image acquisition sequence, where zero contrast agent dose is administered during the diagnostic imaging when these images are acquired.
  • images with other sequences may also be used. For example, in one embodiment, there are five images with five different sequences acquired.
  • the training uses a reference ground truth Tlw image acquired with full contrast agent dose.
  • the images are pre-processed to co-register and normalize them to adjust for acquisition and scaling differences.
  • the normalization may include normalizing the intensity within each contrast via histogram matching on median.
  • Pre-processing may also include bias field correction to correct the bias field distortion via N4ITK algorithm.
  • Pre-processing may also include using a trained classifier (VGG16) to filter out more abnormal slices.
  • the deep learning network is preferably a 2.5D deep convolutional adversarial network. Using the trained network, a first image and second image acquired with different sequences are then used as input to the deep learning network (DLN) which then generates as output a synthesized full-dose contrast agent image of the subject.
  • DLN deep learning network

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Radiology & Medical Imaging (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Public Health (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Quality & Reliability (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Ultra Sonic Daignosis Equipment (AREA)

Abstract

A method for diagnostic imaging with reduced contrast agent dose uses a deep learning network (DLN) [114] that has been trained using zero-contrast [100] and low-contrast [102] images as input to the DLN and full-contrast images [104] as reference ground truth images. Prior to training, the images are pre-processed [106, 110, 118] to co-register and normalize them. The trained DLN [114] is then used to predict a synthesized full-dose contrast agent image [116] from acquired zero-dose and low-dose images.

Description

CONTRAST DOSE REDUCTION FOR MEDICAL IMAGING
USING DEEP LEARNING FIELD OF THE INVENTION
This invention relates generally to medical diagnostic imaging. More specifically, it relates to imaging techniques that use contrast agents.
BACKGROUND OF THE INVENTION
Many types of medical imaging use contrast agents to enhance the visualization of normal and abnormal structures. Examples include conventional angiography, fluoroscopy, computed tomography (CT), ultrasound, and magnetic resonance imaging (MRI). It is often desirable to reduce the contrast agent dose in order to increase the safety of these agents. However, reduced dose reduces desired imaging enhancements, so this has not been possible.
For example, MRI is a powerful imaging technique providing unique information to distinguish different soft tissues and pathologies. Magnetic contrast agents with unique relaxation parameters are often administered to further boost the visibility of pathology and delineation of lesions. Gadolinium based contrast agents (GBCAs) are widely used in MRI exams because of their paramagnetic properties, for applications such as angiography, neuroimaging and liver imaging. However, although GBCAs are designed to avoid toxic release and have been widely applied in contrast enhanced MRI (CE-MRI) examinations to assist the diagnosis, there are several side effects resulting from GBCA administration. Therefore, there are many reasons why it would be advantageous to reduce the dosage of GBCA while preserving the enhanced contrast in MRI imaging. Similar problems are associated with administering contrast agents in other imaging modalities as well. Thus, it would be of benefit to be able to reduce the dose of contrast agents generally in diagnostic imaging techniques, without sacrificing the image enhancement benefits that the contrast agents provide. BRIEF SUMMARY OF THE INVENTION
The present invention provides techniques to enhance image quality of diagnostic imaging modalities using a lower dose of contrast than is currently possible. This enables new opportunities for improving the value of medical imaging. In addition to MRI, the techniques are generally applicable to a variety of diagnostic imaging techniques including angiography, fluoroscopy, computed tomography (CT), and ultrasound.
Surprisingly, the techniques of the present invention are able to predict a synthesized full-dose contrast agent image from a low-dose contrast agent image and a pre-dose image. The low dose may be any fraction of the full dose, but is preferably 1/10 or less of the full dose. Significantly, naively amplifying the contrast enhancement of a 1/10 low-dose CE-MRI by a factor of ten results in poor image quality with widespread noise and ambiguous structures. Even though the low-contrast image cannot be used for diagnosis directly, or by simply amplifying its uptake, the techniques of the present invention remarkably are able to recover the full contrast signal and generate predicted full-contrast images with high diagnostic quality. Specifically, in experimental tests, the method yielded significant improvements over the 10% low-dose images, with over 5 dB PSNR gains, 11% SSIM increased, and improvements in ratings on image quality and contrast enhancement. Embodiments of the invention use a deep learning network, such as a Convolutional Neural Network for image-to-image regression, with a pre-contrast image and low-contrast image as input, and with a predicted full-contrast image as output. A residual learning approach is preferably used in prediction. The method includes preprocessing to co-register and normalize between different images so they are directly comparable. This step is important since there are arbitrary different acquisition and scaling factor for each scan. Preferably, the average signal is used for normalization. The preprocessed images are then used to train the deep learning network to predict the full-contrast image from the pre-contrast and low-contrast images. The trained network is then used to synthesize full-contrast images from clinical scans of pre-contrast and low-contrast images. The techniques are generally applicable for any diagnostic imaging modality that uses a contrast agent. For example, imaging applications would include fluoroscopy, MRI, CT, and ultrasound.
In one aspect, the invention provides a method for training a diagnostic imaging device to perform medical diagnostic imaging with reduced contrast agent dose. The method includes a) performing diagnostic imaging of a set of subjects to produce a set of images comprising, for each subject of the set of subjects, i) a full-contrast image acquired with a full contrast agent dose administered to the subject, ii) a low-contrast image acquired with a low contrast agent dose administered to the subject, where the low contrast agent dose is less than the full contrast agent dose, and iii) a zero-contrast image acquired with no contrast agent dose administered to the subject; b) pre-processing the set of images to co-register and normalize the set of images to adjust for acquisition and scaling differences between different scans; and c) training a deep learning network (DLN) with the pre-processed set of images by applying zero-contrast images from the set of images and low-contrast images from the set of images as input to the DLN and using a cost function to compare the output of the DLN with full-contrast images from the set of images to train parameters of the DLN using backpropagation. The cost function may be, for example, an MAE loss function, or a mixture loss function with non-local structural similarities.
In another aspect, the invention provides a method for medical diagnostic imaging with reduced contrast agent dose. The method includes a) performing diagnostic imaging of a subject to produce a low-contrast image acquired with a low contrast agent dose administered to the subject, where the low contrast agent dose is less than a full contrast agent dose, and a zero- contrast image acquired with no contrast agent dose administered to the subject; b) preprocessing the low-contrast image and zero-contrast image to co-register and normalize the images to adjust for acquisition and scaling differences; and c) applying the low-contrast image and the zero-contrast image as input to a deep learning network (DLN) to generate as output of the DLN a synthesized full-dose contrast agent image of the subject; where the DLN has been trained by applying zero-contrast images and low-contrast images as input and full-contrast images as reference ground-truth images.
The low contrast agent dose is preferably less than 10% of a full contrast agent dose. The diagnostic imaging may be angiography, fluoroscopy, computed tomography (CT), ultrasound, or magnetic resonance imaging. For example, performing diagnostic imaging may include performing magnetic resonance imaging where the full contrast agent dose is at most 0.1 mmol/kg Gadolinium MRI contrast. In one embodiment, the DLN is an encoder-decoder convolutional neural network (CNN) including bypass concatenate connections and residual connections.
In another aspect, the invention provides a method for medical diagnostic imaging including a) performing diagnostic imaging of a subject to produce a first image acquired with a first image acquisition sequence and a second image acquired with a second image acquisition sequence distinct from the first image acquisition sequence, where zero contrast agent dose is administered during the diagnostic imaging; b) pre-processing the first image and the second image to co-register and normalize the images to adjust for acquisition and scaling differences; c) applying the first image and the second image as input to a deep learning network (DLN) to generate as output of the DLN a synthesized full-dose contrast agent image of the subject; wherein the DLN has been trained by applying zero-contrast images with different imaging sequences as input and full-contrast images acquired with a full contrast agent dose as reference ground-truth images.
BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS FIG. 1 is a flow chart showing a processing pipeline according to an embodiment of the invention.
FIG. 2 illustrates a workflow of a protocol and procedure for acquisition of images used for training according to an embodiment of the invention.
FIG. 3 is a schematic overview of an image processing flow according to an embodiment of the invention.
FIG. 4 is an illustration of a signal model to synthesize full-dose contrast-enhanced MRI image according to an embodiment of the invention.
FIG. 5 shows a detailed deep learning (DL) model architecture according to an embodiment of the invention. FIG. 6 is a set of images of a patient with intracranial metastatic disease, showing the predicted image synthesized from pre-contrast image and low-dose image, compared with a full-dose image according to an embodiment of the invention.
FIG. 7 is a set of images of a patient with a brain neoplasm, showing a synthesized full-dose image predicted from a pre-contrast image and low-dose image, compared with a full-dose image according to an embodiment of the invention.
FIG. 8 is a set of images of a patient with a programmable shunt and intracranial hypotension, showing noise suppression capability and consistent contrast enhancement capability according to an embodiment of the invention.
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention provide a deep learning based diagnostic imaging technique to significantly reduce contrast agent dose levels while maintaining diagnostic quality for clinical images.
Detailed illustrations of the protocol and procedure of an embodiment of the invention are shown in FIG. 1 and FIG. 2. Although the embodiments described below focus on MRI imaging for purposes of illustration, the principles and techniques of the invention described herein are not limited to MRI but are generally applicable to various imaging modalities that make use of contrast agents.
FIG. 1 is a flow chart showing a processing pipeline for an embodiment of the invention. A deep learning network is trained using multi -contrast images 100, 102, 104 acquired from scans of a multitude of subjects with a wide range of clinical indications. The images are pre- processed to perform image co-registration 106, to produce multi -contrast images 108, and data augmentation 110 to produce normalized multi -contrast image patches 112. Rigid or non-rigid co-registration may be used to adjust multiple image slices or volumes to match the pixels and voxels to each other. Since there can be arbitrary scaling differences between different volumes, normalization is used to match the intensity of each image/volume. Brain and anatomy masks are used optionally to extract the important regions of interests in each image/volume. Reference images 104 are also processed to perform co-registration and normalization 118. These pre-processed images are then used to train a deep learning network 114, which is preferably implemented using residual learning in a convolutional neural network. The input to the deep learning network is a zero-contrast dose image 100 and low-contrast dose image 102, while the output of the network is a synthesized prediction of a full-contrast dose image 116. During training, a reference full contrast image 104 is compared with the synthesized image 116 using a loss function to train the network using error backpropagation.
FIG. 2 illustrates the workflow of the protocol and procedure for acquisition of images used for training. After a pre-contrast (zero-dose) image 200 is acquired, a low dose (e.g., 10%) of contrast is administered and a low-dose image 202 is acquired. An additional dose (e.g., 90%) of contrast is then administered to total a full 100% dose, and a full-dose image 204 is then acquired.
In an MRI embodiment, images are acquired with 3T MRI scanners (GE Healthcare, Waukesha, WI, USA) using standard neuro clinical protocol with high-resolution 3D Tl -weighted inversion-recovery prepped fast-spoiled-gradient-echo (IR-FSPGR) imaging at 3T. Specifically, high-resolution Tl -weighted IR-FSPGR pre-contrast images, post-contrast images with 10% low-dose and 100% full-dose of gadobenate dimeglumine (0.01 and 0.1 mmol/kg, respectively) full-dose images are acquired.
For other imaging modalities, similar setups are used to include at least one set of images without enhancement that is acquired without injecting contrasts. And optionally at least one set of images with low-level enhancement that is acquired with injecting low dosage of contrasts. For CT, the contrast is usually an iodine-based contrast agent. The CT contrast agent is usually administered based on preference of physicians, regulatory standard and patient weight. The administration of a low dose means injecting less than standard protocols. For CT there can be multiple set of images with different contrast visualization (possibly the same dosage but different visual appearance) by using multiple energy radiation. For CT, the ground-truth is a set of images acquired with 100% full-dose of CT contrast. For ultrasound, the contrast agent can be, for example, microbubbles. Similarly, the scans may contain at least one image with certain dosage of contrast and optionally images with variable dosage or images with dosage shown differently as acquired with different probe frequency. FIG. 3 is a schematic overview of the image processing flow. The steps of the detailed workflow are applied to the multi -contrast images acquired for each subject. In acquisition stage 300, scans are performed as described above to acquire a zero-contrast image 306, low-contrast image 308, and full-contrast image 310. These multi -contrast images then pass through a preprocessing stage 302. The resulting images then are used in a deep learning training stage 304 for training a deep learning network to synthesize a full-dose image 312. Pre-processing steps include image co-registration and signal normalization. Normalization is used to remove bias in the images obtained at different dosage levels, which is performed using a mask 314 of non- enhancing tissues. Specifically, to remove the systematic differences between different signal intensity levels in non-enhancing regions (such as scalp fat), co-registration and signal normalization based on average voxel value within a mask is performed. Alternatively, the normalization can be based on max/min intensity of the images or certain percentiles. Also, the normalization can be based on matching the distribution of intensity to certain predefined or standard distribution. Mask can be applied to calculate the average, max, min, percentile or the distribution to get better normalization performance to match different set of images. This step is performed because the transmit and receive gains used for 3 sequences of the 3 different scans are not guaranteed to be the same. For CT and ultrasound the scaling and normalization is also applicable as there are possible intensity re-scaling steps in both acquisition and image storage process.
After co-registration and normalization, the remaining signal differences between different contrast level images is in theory only related to contrast uptakes and non-structural image noise, as shown in FIG. 4, which is an illustration of the signal model to synthesize full-dose contrast-enhanced MRI image 412. Compared with the reference high-quality contrast uptake 406 between pre-contrast 400 and full-dose 404 CE-MRI, the low-dose contrast uptake 408 between pre-contrast 400 and low-dose 402 is noisy but does include contrast information. By using a deep learning method, we learn the denoising to generate high-quality predicted contrast uptake 410 and then combine this with the pre-contrast scan 400 to synthesize a full-dose CE- MRI image 412. After pre-processing, a deep learning network is trained using the true 100% full-dose CE-MRI images as the reference ground-truth. The non-contrast (zero-dose) MRI and the 10% low-dose CE-MRI are provided to the network as inputs, and the output of the network is an approximation of the full-dose CE-MRI. During training, this network implicitly learns the guided denoising of the noisy contrast uptake extracted from the difference signal between low- dose and non-contrast (zero-dose) images, which can be scaled to generate the contrast enhancement of a full-dose image.
The detailed deep learning (DL) model architecture in one embodiment of the invention is shown in FIG. 5. This model is an encoder-decoder convolutional neural network with 3 encoder steps 500, 502, 504 and 3 decoder steps 506, 508, 510. In each step, there are 3 convolutional layers connected by 3x3 Conv-BN-ReLU. Encoding steps are connected in sequence by 2x2 max-pooling, and decoder steps are connected in sequence by 2x2 up- sampling. Bypass concatenate connections 512, 514, 516 combine symmetric layers to avoid resolution loss. The residual connections 518, 520, 522, 524, 526, 528, 530 enable the model to synthesize a full-dose image by predicting the enhancement signal 540 from a difference between pre-dose image 542 and low-dose image 544. The cost function 532 compares the predicted full-dose image 534 and the reference ground-truth full-dose image 536, which enables the optimization of the network parameters via error backpropagation 538.
The above-described network architecture is just one illustrative example. Other architectures are possible. For example, the network can have different number of layers, image size in each layers and variable connections between layers. The function in each layer can be different linear or nonlinear functions. The output layer can have a different so-called activation function that maps to certain range of output intensity. There can be multiple number of networks concatenated, so called recurrent network, that further improves the capability that one single network can achieve.
In one test, the network was trained on around 300 2D slices of the co-registered 3D volumes in each patient, excluding the slices at the base of the brain which had low S R and no valuable information for anatomy or contrast. Standard image rigid transformations are used to further augment the dataset in training to ensure the robustness of the model. In training, stochastic gradient descent (SGD) was used for each subset of training datasets and backpropagation is used to optimize the network parameters with respect to a cost function comparing predicted and true full-dose MR images. In one embodiment, the mean-absolute- error (MAE) cost function, also known as the LI loss, is used in training. Training takes 200 epochs with SGD and ADAM method for optimization and 10% of the training dataset with random permutations was used for validation to optimize hyper-parameters and pick out the best model among all iterations. SGD and ADAM are also examples of operators to solve the optimization problem in training the network. There are many other options including, for example, RMSprop and Adagrad. Basically, the optimizers enable faster and smoother convergence.
The loss functions can include, for example, a function map from the ground-truth image and predicted image pair to a set loss values. This includes pixel-wise or voxel-wise loss that is based on pixel/voxel differences which usually use LI (mean-absolute-error) or L2 (mean- squared-error) loss. Also, the loss can be based on regions with certain size that considering similarity of the structures, e.g., SSIM structural similarities. Or the loss can be computed based on other previously or concurrently trained networks that so-called perceptual loss or adversarial loss respectively. Also the loss can be any arbitrary weighted combination of many loss functions.
In a clinical setting, the trained deep learning network is used to synthesize a full-dose image from zero-dose and low-dose images. The co-registered and normalized non-contrast (zero- dose) and low-dose images are loaded from DICOM files and input to the trained network. With efficient forward-passing, which takes around 0.1 sec per 512-by-512 image, the synthesized full-dose image is generated. The process is then conducted for each 2D slice to generate entire 3D volume and stored in a new DICOM folder for further evaluation. For both training and testing the neural network, we used the Keras framework with Tensorflow backend, CUDA8 and CUDNN5.1, on a Linux server with 2 NVIDIA GTX 1080-TI GPUs. This is but one illustrative example. There are many alternative software and hardware implementations of deep networks that may be used. In tests of the technique, the synthesized full-dose CE-MRI images contain consistent contrast uptake and similar enhancement. Detailed comparisons for several representative cases are discussed below, demonstrating that the method can generate synthesized full-dose contrast- enhanced MRI with consistent enhancement but using lower dose (e.g., 1/10 full dose).
As shown in FIG. 6, in a patient with intracranial metastatic disease, the predicted image 606 synthesized from pre-contrast image 600 and low-dose image 602 has similar highlighting of contrast enhancement in the lesions as full-dose image 604. The lesions show improved visibility in the synthesized full-contrast image 606 while they cannot be reliably appreciated in low-dose CE-MRI image 602. Moreover, the synthesized CE-MRI image 606 shows a similar outline of a metastatic lesion in the right posterior cerebellum compared with the one in true full-contrast CE-MRI image 604.
For another case in a patient with a brain neoplasm, shown in FIG. 7, the synthesized full-dose CE-MRI image 706 predicted from the pre-contrast image 700 and low-dose image 702 shows similar contrast uptake as in the real full-dose CE-MRI image 704. Plus, the image quality of the synthesized full-dose image 706 is better compared with the acquired full-dose CE-MRI 704 which has more severe motion artifacts and aliasing. Surprisingly, this demonstrates that the synthesized results have additional advantages over the original full-dose CE-MRI for better denoising and artifact removal, e.g., suppression of motion artifacts and aliasing.
In a case of a patient with a programmable shunt and intracranial hypotension, shown in FIG. 8, results also demonstrate better noise suppression capability and consistent contrast enhancement capability. The engorgement of the dural sinuses and pachymeningeal enhancement are clearly seen on both the synthesized image 806, predicted from pre-contrast image 800 and low-dose image 802, and true full-dose image 804, while there is less noise in non-enhancing regions on the synthesized image. Thus, the method generates diagnostic quality contrast from the low-dose acquisition, also demonstrating improved noise suppression and contrast enhancement.
Tests of the method by the inventors have demonstrated that the synthesized CE-MRI images are improved significantly (both p < 0.001) compared with the low-dose image on all quantitative metrics with 11.0% improvements in SSIM and over 5.0 dB gains in PS R. Compared with the true acquired full-dose CE-MRI, the proposed method is not significantly different (both p > 0.05) for overall image quality and the clarity of the enhanced contrast. Comparably, the 10% low-contrast images are significantly (p < 0.001) worse in the clarity of the enhanced contrast, with a decrease of over 3 on a 5-point scale.
As FIG. 7 and FIG. 8 show, and as verified by neuroradiologists' ratings, the synthesized full- dose images show significant (p < 0.05) improvements over the acquired full-dose CE-MRI in suppressing artifacts in the non-enhancing regions, which was an additional advantage of the DL method.
Compared with real full-dose CE-MRI, no significant differences (p>0.05) in contrast enhancements information could be identified by neuro-radiologists blinded to image origin. The synthesized full-dose images also appear to better suppress image artifacts (p<0.05). This is because the high-intensity signal of the real full-dose enhanced contrast inevitably augments motion/aliasing artifacts which are better suppressed when estimating contrast from the low- intensity contrast enhancements. The impact of being able to reduce contrast dose by 90% while retaining diagnostic information could have a large impact on patient well-being and imaging costs.
The illustrative embodiment described above uses a 2D CNN. Further performance gains are achievable with 3D CNN models considering the correlated spatial information in adjacent thin slices. Thus, the implementation of the invention illustrated with a 2D CNN is only an illustration, and 3D implementations are envisioned within the scope of the invention. More generally, 2D (process single slice of image), 2.5D (multiple slices of image), 3D (entire volume) and patch based (crop from volume) algorithms are envisioned.
In addition, the inventors envision that with more datasets, more complicated but powerful models such as generative models can be trained to further improve the contrast synthesis. Finally, although the MAE loss is used as cost function to train the DL model in the above illustrative embodiment, other loss functions are envisioned, such as mixture loss functions with non-local structural similarities or using Generative Adversarial Network (GAN) as a data driven cost function.
In the illustrative embodiment described above, a fixed level of gadolinium contrast is used with 90% reduction from the original clinical usage level.
More generally, the term full-dose is used herein to refer to a standard dosage enabling certain defined visual quality. Usually the contrasts are treated as drug and they have FDA regulated full-dose that should be used. The full-dose means a standard dosage that is recommended or required by the FDA or clinicians to achieve good diagnostic quality.
However, it is envisioned that additional reduction of the low dose is feasible by using alternative network structures and loss function, using larger training datasets with various pathologies as well as other complimentary information such as multi -contrast MRI, multiple imaging modalities, and patients' clinical history information. It is also envisioned that the technique can make use of a patch based method wherein the input image size can be different than 128x128x1. It can be any xxyxz size that uses a patch of size x*y and considering depth/thickness of z. It is also envisioned that low dose levels higher than 10% may be used. In general, it is beneficial to minimize the dose subject to the constraint that the synthesized image retains diagnostic quality. The resulting minimal low dose may vary depending on contrast agent and imaging modality. It is also possible to base the training on more than two images having different dose levels, and it is not necessary to include a zero-dose image. In general, if the training dataset has a sufficient number of paired scans with two distinct dose levels, the network can be trained to automatically generalized to improve the lower dose
Figure imgf000014_0001
to the higher dose i¾igh, where the higher dose level is a full dose or less, and the lower dose level is less than the high dose, i.e., 0 <
Figure imgf000014_0002
< dhigh≤ 1. The inventors also envision synthesis of high dose images from zero dose images, without the need for any contrast dose images. The zero dose images in this case include multiple MR images acquired using different sequences (e.g., Tlw, T2w, FLAIR, DWI) that show different appearance/intensity of different tissues. The trained network can then predict a synthesized full- contrast Tlw to be used for diagnosis. Thus, according to this method, the deep learning network is trained with a first image acquired with a first image acquisition sequence and a second image acquired with a second image acquisition sequence distinct from the first image acquisition sequence, where zero contrast agent dose is administered during the diagnostic imaging when these images are acquired. In addition, images with other sequences may also be used. For example, in one embodiment, there are five images with five different sequences acquired. The training uses a reference ground truth Tlw image acquired with full contrast agent dose. The images are pre-processed to co-register and normalize them to adjust for acquisition and scaling differences. The normalization may include normalizing the intensity within each contrast via histogram matching on median. Pre-processing may also include bias field correction to correct the bias field distortion via N4ITK algorithm. Pre-processing may also include using a trained classifier (VGG16) to filter out more abnormal slices. The deep learning network is preferably a 2.5D deep convolutional adversarial network. Using the trained network, a first image and second image acquired with different sequences are then used as input to the deep learning network (DLN) which then generates as output a synthesized full-dose contrast agent image of the subject.

Claims

1. A method for training a diagnostic imaging device to perform medical diagnostic imaging with reduced contrast agent dose, the method comprising:
a) performing diagnostic imaging of a set of subjects to produce a set of images comprising, for each subject of the set of subjects, i) a full-contrast image acquired with a full contrast agent dose administered to the subject, ii) a low-contrast image acquired with a low contrast agent dose administered to the subject, where the low contrast agent dose is less than the full contrast agent dose, and iii) a zero-contrast image acquired with no contrast agent dose administered to the subject;
b) pre-processing the set of images to co-register and normalize the set of images to adjust for acquisition and scaling differences between different scans;
c) training a deep learning network (DLN) with the pre-processed set of images by applying zero-contrast images from the set of images and low-contrast images from the set of images as input to the DLN and using a cost function to compare the output of the DLN with full-contrast images from the set of images to train parameters of the DLN using backpropagation.
2. The method of claim 1 wherein the cost function is an MAE loss function, or a mixture loss function with non-local structural similarities.
3. A method for medical diagnostic imaging with reduced contrast agent dose, the method comprising:
a) performing diagnostic imaging of a subject to produce a low-contrast image acquired with a low contrast agent dose administered to the subject, where the low contrast agent dose is less than a full contrast agent dose, and a zero-contrast image acquired with no contrast agent dose administered to the subject;
b) pre-processing the low-contrast image and zero-contrast image to co-register and normalize the images to adjust for acquisition and scaling differences;
c) applying the low-contrast image and the zero-contrast image as input to a deep learning network (DLN) to generate as output of the DLN a synthesized full-dose contrast agent image of the subject; wherein the DLN has been trained by applying zero-contrast images and low-contrast images as input and full-contrast images as reference ground-truth images.
4. The method of claim 3 wherein
the low contrast agent dose is less than 10% of a full contrast agent dose.
5. The method of claim 3 wherein
performing diagnostic imaging comprises performing angiography, fluoroscopy, computed tomography (CT), ultrasound, or magnetic resonance imaging.
6. The method of claim 3 wherein
performing diagnostic imaging comprises performing magnetic resonance imaging, and the full contrast agent dose is at most 0.1 mmol/kg Gadolinium MRI contrast.
7. The method of claim 3 wherein
the DLN is an encoder-decoder convolutional neural network (CNN) including bypass concatenate connections and residual connections.
8. A method for medical diagnostic imaging, the method comprising:
a) performing diagnostic imaging of a subject to produce a first image acquired with a first image acquisition sequence and a second image acquired with a second image acquisition sequence distinct from the first image acquisition sequence, where zero contrast agent dose is administered during the diagnostic imaging;
b) pre-processing the first image and the second image to co-register and normalize the images to adjust for acquisition and scaling differences;
c) applying the first image and the second image as input to a deep learning network (DLN) to generate as output of the DLN a synthesized full-dose contrast agent image of the subject;
wherein the DLN has been trained by applying zero-contrast images with different imaging sequences as input and full-contrast images acquired with a full contrast agent dose as reference ground-truth images.
PCT/US2018/055034 2017-10-09 2018-10-09 Contrast dose reduction for medical imaging using deep learning WO2019074938A1 (en)

Priority Applications (9)

Application Number Priority Date Filing Date Title
AU2018346938A AU2018346938B2 (en) 2017-10-09 2018-10-09 Contrast dose reduction for medical imaging using deep learning
CA3078728A CA3078728A1 (en) 2017-10-09 2018-10-09 Contrast dose reduction for medical imaging using deep learning
CN201880072487.0A CN111601550B (en) 2017-10-09 2018-10-09 Contrast agent reduction for medical imaging using deep learning
JP2020520027A JP7244499B2 (en) 2017-10-09 2018-10-09 Contrast Agent Dose Reduction in Medical Imaging Using Deep Learning
KR1020207013155A KR20200063222A (en) 2017-10-09 2018-10-09 Contrast dose reduction in medical imaging with deep learning
BR112020007105-6A BR112020007105A2 (en) 2017-10-09 2018-10-09 method for training a diagnostic imaging device to perform a medical diagnostic imaging with a reduced dose of contrast agent
EP18866524.4A EP3694413A4 (en) 2017-10-09 2018-10-09 Contrast dose reduction for medical imaging using deep learning
SG11202003232VA SG11202003232VA (en) 2017-10-09 2018-10-09 Contrast dose reduction for medical imaging using deep learning
JP2023036252A JP7476382B2 (en) 2017-10-09 2023-03-09 Contrast agent dose reduction in medical imaging using deep learning

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US201762570068P 2017-10-09 2017-10-09
US62/570,068 2017-10-09

Publications (1)

Publication Number Publication Date
WO2019074938A1 true WO2019074938A1 (en) 2019-04-18

Family

ID=65994007

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2018/055034 WO2019074938A1 (en) 2017-10-09 2018-10-09 Contrast dose reduction for medical imaging using deep learning

Country Status (10)

Country Link
US (2) US10997716B2 (en)
EP (1) EP3694413A4 (en)
JP (2) JP7244499B2 (en)
KR (1) KR20200063222A (en)
CN (1) CN111601550B (en)
AU (1) AU2018346938B2 (en)
BR (1) BR112020007105A2 (en)
CA (1) CA3078728A1 (en)
SG (1) SG11202003232VA (en)
WO (1) WO2019074938A1 (en)

Cited By (23)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021044153A1 (en) * 2019-09-04 2021-03-11 Oxford University Innovation Limited Enhancement of medical images
WO2021052896A1 (en) 2019-09-18 2021-03-25 Bayer Aktiengesellschaft Forecast of mri images by means of a forecast model trained by supervised learning
WO2021052850A1 (en) 2019-09-18 2021-03-25 Bayer Aktiengesellschaft Generation of mri images of the liver
WO2021069343A1 (en) 2019-10-11 2021-04-15 Bayer Aktiengesellschaft Acceleration of mri examinations
WO2022106302A1 (en) 2020-11-20 2022-05-27 Bayer Aktiengesellschaft Representation learning
WO2022179896A2 (en) 2021-02-26 2022-09-01 Bayer Aktiengesellschaft Actor-critic approach for generating synthetic images
WO2022184297A1 (en) 2021-03-02 2022-09-09 Bayer Aktiengesellschaft Machine learning in the field of contrast-enhanced radiology
WO2022189015A1 (en) 2021-03-09 2022-09-15 Bayer Aktiengesellschaft Machine learning in the field of contrast-enhanced radiology
WO2022223383A1 (en) 2021-04-21 2022-10-27 Bayer Aktiengesellschaft Implicit registration for improving synthesized full-contrast image prediction tool
JP2022550688A (en) * 2019-09-25 2022-12-05 サトゥル メディカル，インコーポレイテッド Systems and methods for improving low-dose volume-enhanced MRI
EP4174868A1 (en) 2021-11-01 2023-05-03 Bayer Aktiengesellschaft Synthetic contrast-enhanced mr images
WO2023073165A1 (en) 2021-11-01 2023-05-04 Bayer Aktiengesellschaft Synthetic contrast-enhanced mr images
EP4202855A1 (en) 2021-12-22 2023-06-28 Guerbet Method for processing at least a pre-contrast image and a contrast image respectively depicting a body part prior to and after an injection of a dose of contrast agent
EP4210069A1 (en) 2022-01-11 2023-07-12 Bayer Aktiengesellschaft Synthetic contrast-enhanced ct images
EP4233726A1 (en) 2022-02-24 2023-08-30 Bayer AG Prediction of a representation of an area of an object to be examined after the application of different amounts of a contrast agent
US11915361B2 (en) 2019-09-18 2024-02-27 Bayer Aktiengesellschaft System, method, and computer program product for predicting, anticipating, and/or assessing tissue characteristics
EP4332601A1 (en) 2022-09-05 2024-03-06 Bayer AG Generation of artificial contrast agent-enhanced radiological recordings
WO2024046833A1 (en) 2022-08-30 2024-03-07 Bayer Aktiengesellschaft Generation of synthetic radiological images
WO2024046831A1 (en) 2022-08-30 2024-03-07 Bayer Aktiengesellschaft Generation of synthetic radiological images
EP4336513A1 (en) 2022-08-30 2024-03-13 Bayer Aktiengesellschaft Generation of synthetic radiological recordings
WO2024052156A1 (en) 2022-09-05 2024-03-14 Bayer Aktiengesellschaft Generation of artificial contrast-enhanced radiological images
EP4369353A1 (en) 2022-11-12 2024-05-15 Bayer Aktiengesellschaft Generation of artificial contrast agent-enhanced radiological recordings
EP4369285A1 (en) 2022-11-12 2024-05-15 Bayer AG Generation of artificial contrast agent-enhanced radiological recordings

Families Citing this family (41)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11100621B2 (en) * 2017-10-20 2021-08-24 Imaging Biometrics, Llc Simulated post-contrast T1-weighted magnetic resonance imaging
EP3477583A1 (en) * 2017-10-31 2019-05-01 Koninklijke Philips N.V. Deep-learning based processing of motion artifacts in magnetic resonance imaging data
US11880962B2 (en) * 2018-02-15 2024-01-23 General Electric Company System and method for synthesizing magnetic resonance images
US11896360B2 (en) * 2018-03-12 2024-02-13 Lvis Corporation Systems and methods for generating thin image slices from thick image slices
US11164067B2 (en) * 2018-08-29 2021-11-02 Arizona Board Of Regents On Behalf Of Arizona State University Systems, methods, and apparatuses for implementing a multi-resolution neural network for use with imaging intensive applications including medical imaging
US20210358183A1 (en) * 2018-09-28 2021-11-18 Mayo Foundation For Medical Education And Research Systems and Methods for Multi-Kernel Synthesis and Kernel Conversion in Medical Imaging
US20200294288A1 (en) * 2019-03-13 2020-09-17 The Uab Research Foundation Systems and methods of computed tomography image reconstruction
CN110223255B (en) * 2019-06-11 2023-03-14 太原科技大学 Low-dose CT image denoising and recursion method based on residual error coding and decoding network
US20210150671A1 (en) * 2019-08-23 2021-05-20 The Trustees Of Columbia University In The City Of New York System, method and computer-accessible medium for the reduction of the dosage of gd-based contrast agent in magnetic resonance imaging
JP2021036969A (en) * 2019-08-30 2021-03-11 キヤノン株式会社 Machine learning device, machine learning method, and program
WO2021061710A1 (en) 2019-09-25 2021-04-01 Subtle Medical, Inc. Systems and methods for improving low dose volumetric contrast-enhanced mri
CN110852993B (en) * 2019-10-12 2024-03-08 拜耳股份有限公司 Imaging method and device under action of contrast agent
JP7394588B2 (en) * 2019-11-07 2023-12-08 キヤノン株式会社 Information processing device, information processing method, and imaging system
US10984530B1 (en) * 2019-12-11 2021-04-20 Ping An Technology (Shenzhen) Co., Ltd. Enhanced medical images processing method and computing device
EP3872754A1 (en) * 2020-02-28 2021-09-01 Siemens Healthcare GmbH Method and system for automated processing of images when using a contrast agent in mri
CN111325695B (en) * 2020-02-29 2023-04-07 深圳先进技术研究院 Low-dose image enhancement method and system based on multi-dose grade and storage medium
CN111388000B (en) * 2020-03-27 2023-08-25 上海杏脉信息科技有限公司 Virtual lung air retention image prediction method and system, storage medium and terminal
CN112508835B (en) * 2020-12-10 2024-04-26 深圳先进技术研究院 GAN-based contrast agent-free medical image enhancement modeling method
WO2022120734A1 (en) * 2020-12-10 2022-06-16 深圳先进技术研究院 Contrast-agent-free medical image enhancement method based on gan
CN112634390B (en) * 2020-12-17 2023-06-13 深圳先进技术研究院 High-energy image synthesis method and device for generating countermeasure network model based on Wasserstein
CN112733624B (en) * 2020-12-26 2023-02-03 电子科技大学 People stream density detection method, system storage medium and terminal for indoor dense scene
KR102316312B1 (en) * 2021-02-01 2021-10-22 주식회사 클라리파이 Apparatus and method for contrast amplification of contrast-enhanced ct images based on deep learning
JP7376053B2 (en) 2021-02-01 2023-11-08 クラリピーアイ インコーポレイテッド Contrast enhancement CT image contrast amplification device and method based on deep learning
US20240065772A1 (en) * 2021-02-03 2024-02-29 Cordiguide Ltd. Navigation assistance in a medical procedure
EP4044120A1 (en) * 2021-02-15 2022-08-17 Koninklijke Philips N.V. Training data synthesizer for contrast enhancing machine learning systems
US11727087B2 (en) * 2021-04-05 2023-08-15 Nano-X Ai Ltd. Identification of a contrast phase depicted in a medical image
US11967066B2 (en) * 2021-04-12 2024-04-23 Daegu Gyeongbuk Institute Of Science And Technology Method and apparatus for processing image
CN113033704B (en) * 2021-04-22 2023-11-07 江西理工大学 Intelligent judging method and system for copper converter converting copper-making final point based on pattern recognition
CN113379616B (en) * 2021-04-28 2023-09-05 北京航空航天大学 Method for generating gadolinium contrast agent enhanced magnetic resonance image
US20220381861A1 (en) * 2021-05-19 2022-12-01 Siemens Healthcare Gmbh Method and system for accelerated acquisition and artifact reduction of undersampled mri using a deep learning based 3d generative adversarial network
EP4095796A1 (en) * 2021-05-29 2022-11-30 Bayer AG Machine learning in the field of radiology with contrast agent
EP4113537A1 (en) * 2021-06-30 2023-01-04 Guerbet Methods for training a prediction model, or for processing at least a pre-contrast image depicting a body part prior to an injection of contrast agent using said prediction model
US11903760B2 (en) * 2021-09-08 2024-02-20 GE Precision Healthcare LLC Systems and methods for scan plane prediction in ultrasound images
WO2023055721A1 (en) * 2021-09-29 2023-04-06 Subtle Medical, Inc. Systems and methods for contrast dose reduction
EP4334900A1 (en) * 2021-10-15 2024-03-13 Bracco Imaging S.p.A. Training a machine learning model for simulating images at higher dose of contrast agent in medical imaging applications
EP4224420A1 (en) 2022-02-04 2023-08-09 Siemens Healthcare GmbH A computer-implemented method for determining scar segmentation
EP4246526A1 (en) * 2022-03-17 2023-09-20 Koninklijke Philips N.V. System and method for providing enhancing or contrast agent advisability indicator
WO2023192392A1 (en) * 2022-03-29 2023-10-05 Angiowave Imaging, Llc System and method for angiographic dose reduction using machine learning
EP4290263A1 (en) * 2022-06-09 2023-12-13 Rheinische Friedrich-Wilhelms-Universität Bonn Method and apparatus for providing mri images relating to at least one body portion of a patient with reduced administration of contrast agent
WO2024044476A1 (en) * 2022-08-23 2024-02-29 Subtle Medical, Inc. Systems and methods for contrast-enhanced mri
WO2024046832A1 (en) * 2022-08-30 2024-03-07 Bayer Aktiengesellschaft Generation of synthetic radiological images

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050171409A1 (en) * 2004-01-30 2005-08-04 University Of Chicago Automated method and system for the detection of lung nodules in low-dose CT image for lung-cancer screening
US20150201895A1 (en) * 2012-08-31 2015-07-23 The University Of Chicago Supervised machine learning technique for reduction of radiation dose in computed tomography imaging
US20160106321A1 (en) * 2013-10-17 2016-04-21 Siemens Aktiengesellschaft Method and System for Machine Learning Based Assessment of Fractional Flow Reserve
WO2016175755A1 (en) * 2015-04-28 2016-11-03 Siemens Healthcare Gmbh METHOD AND SYSTEM FOR SYNTHESIZING VIRTUAL HIGH DOSE OR HIGH kV COMPUTED TOMOGRAPHY IMAGES FROM LOW DOSE OR LOW kV COMPUTED TOMOGRAPHY IMAGES
US20180018757A1 (en) * 2016-07-13 2018-01-18 Kenji Suzuki Transforming projection data in tomography by means of machine learning

Family Cites Families (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE10221643A1 (en) 2002-05-15 2003-12-04 Siemens Ag Image data correcting method for MRI tomography apparatus used for e.g. heart perfusion measurement, involves subtracting value of image element in uniform region from other respective values of image element in normalized image data
DE102006001681B4 (en) 2006-01-12 2008-07-10 Wismüller, Axel, Dipl.-Phys. Dr.med. Method and device for displaying multi-channel image data
JP2007264951A (en) 2006-03-28 2007-10-11 Dainippon Printing Co Ltd Medical image correction device
WO2009020687A2 (en) * 2007-05-18 2009-02-12 Henry Ford Health System Mri estimation of contrast agent concentration using a neural network approach
US8737715B2 (en) * 2009-07-13 2014-05-27 H. Lee Moffitt Cancer And Research Institute, Inc. Methods and apparatus for diagnosis and/or prognosis of cancer
BR112012033776A2 (en) * 2010-07-12 2016-11-01 Ge Healthcare As composition and method of x-ray examination.
US9378548B2 (en) * 2011-12-01 2016-06-28 St. Jude Children's Research Hospital T2 spectral analysis for myelin water imaging
CN103823956B (en) * 2012-11-19 2017-10-31 西门子（中国）有限公司 Determine that contrast agent uses method, device and the imaging device of parameter
US20150025666A1 (en) * 2013-07-16 2015-01-22 Children's National Medical Center Three dimensional printed replicas of patient's anatomy for medical applications
US11647915B2 (en) * 2014-04-02 2023-05-16 University Of Virginia Patent Foundation Systems and methods for medical imaging incorporating prior knowledge
US9990712B2 (en) 2015-04-08 2018-06-05 Algotec Systems Ltd. Organ detection and segmentation
EP3295202B1 (en) * 2015-05-08 2018-07-25 Max-Planck-Gesellschaft zur Förderung der Wissenschaften e.V. Method and device for magnetic resonance imaging with improved sensitivity by noise reduction
JP6450053B2 (en) 2015-08-15 2019-01-09 セールスフォース ドット コム インコーポレイティッド Three-dimensional (3D) convolution with 3D batch normalization
WO2017030276A1 (en) 2015-08-17 2017-02-23 삼성전자(주) Medical image display device and medical image processing method
US10547873B2 (en) * 2016-05-23 2020-01-28 Massachusetts Institute Of Technology System and method for providing real-time super-resolution for compressed videos
WO2017223560A1 (en) 2016-06-24 2017-12-28 Rensselaer Polytechnic Institute Tomographic image reconstruction via machine learning
US10096109B1 (en) * 2017-03-31 2018-10-09 The Board Of Trustees Of The Leland Stanford Junior University Quality of medical images using multi-contrast and deep learning
US11100621B2 (en) * 2017-10-20 2021-08-24 Imaging Biometrics, Llc Simulated post-contrast T1-weighted magnetic resonance imaging
US10482600B2 (en) * 2018-01-16 2019-11-19 Siemens Healthcare Gmbh Cross-domain image analysis and cross-domain image synthesis using deep image-to-image networks and adversarial networks
US10789696B2 (en) * 2018-05-24 2020-09-29 Tfi Digital Media Limited Patch selection for neural network based no-reference image quality assessment
US10665011B1 (en) * 2019-05-31 2020-05-26 Adobe Inc. Dynamically estimating lighting parameters for positions within augmented-reality scenes based on global and local features

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050171409A1 (en) * 2004-01-30 2005-08-04 University Of Chicago Automated method and system for the detection of lung nodules in low-dose CT image for lung-cancer screening
US20150201895A1 (en) * 2012-08-31 2015-07-23 The University Of Chicago Supervised machine learning technique for reduction of radiation dose in computed tomography imaging
US20160106321A1 (en) * 2013-10-17 2016-04-21 Siemens Aktiengesellschaft Method and System for Machine Learning Based Assessment of Fractional Flow Reserve
WO2016175755A1 (en) * 2015-04-28 2016-11-03 Siemens Healthcare Gmbh METHOD AND SYSTEM FOR SYNTHESIZING VIRTUAL HIGH DOSE OR HIGH kV COMPUTED TOMOGRAPHY IMAGES FROM LOW DOSE OR LOW kV COMPUTED TOMOGRAPHY IMAGES
US20180018757A1 (en) * 2016-07-13 2018-01-18 Kenji Suzuki Transforming projection data in tomography by means of machine learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
MARDANI, M. ET AL.: "Deep Generative Adversarial Networks for Compressed Sensing (GANCS) Automates MRI", 31 May 2017 (2017-05-31), XP055593025, Retrieved from the Internet <URL:https://arxiv.org/pdf/1706.00051.pdf> [retrieved on 20190206] *
XIANG LEI ET AL.: "Deep auto-context convolutional neural networks for standard-dose PET image estimation from low-dose PET/MRI", NEUROCOMPUTING, vol. 267, pages 406 - 416, XP085154155, DOI: 10.1016/j.neucom.2017.06.048

Cited By (32)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021044153A1 (en) * 2019-09-04 2021-03-11 Oxford University Innovation Limited Enhancement of medical images
EP4231037A1 (en) 2019-09-18 2023-08-23 Bayer Aktiengesellschaft Acceleration of mri examinations
WO2021052896A1 (en) 2019-09-18 2021-03-25 Bayer Aktiengesellschaft Forecast of mri images by means of a forecast model trained by supervised learning
WO2021052850A1 (en) 2019-09-18 2021-03-25 Bayer Aktiengesellschaft Generation of mri images of the liver
US11915361B2 (en) 2019-09-18 2024-02-27 Bayer Aktiengesellschaft System, method, and computer program product for predicting, anticipating, and/or assessing tissue characteristics
US11727571B2 (en) 2019-09-18 2023-08-15 Bayer Aktiengesellschaft Forecast of MRI images by means of a forecast model trained by supervised learning
JP2022550688A (en) * 2019-09-25 2022-12-05 サトゥル メディカル，インコーポレイテッド Systems and methods for improving low-dose volume-enhanced MRI
WO2021069343A1 (en) 2019-10-11 2021-04-15 Bayer Aktiengesellschaft Acceleration of mri examinations
EP4241672A2 (en) 2019-10-11 2023-09-13 Bayer Aktiengesellschaft Acceleration of mri examinations
WO2022106302A1 (en) 2020-11-20 2022-05-27 Bayer Aktiengesellschaft Representation learning
WO2022179896A2 (en) 2021-02-26 2022-09-01 Bayer Aktiengesellschaft Actor-critic approach for generating synthetic images
WO2022184298A1 (en) * 2021-03-02 2022-09-09 Bayer Aktiengesellschaft System, method, and computer program product for contrast-enhanced radiology using machine learning
WO2022184297A1 (en) 2021-03-02 2022-09-09 Bayer Aktiengesellschaft Machine learning in the field of contrast-enhanced radiology
WO2022189015A1 (en) 2021-03-09 2022-09-15 Bayer Aktiengesellschaft Machine learning in the field of contrast-enhanced radiology
WO2022223383A1 (en) 2021-04-21 2022-10-27 Bayer Aktiengesellschaft Implicit registration for improving synthesized full-contrast image prediction tool
WO2023073165A1 (en) 2021-11-01 2023-05-04 Bayer Aktiengesellschaft Synthetic contrast-enhanced mr images
EP4174868A1 (en) 2021-11-01 2023-05-03 Bayer Aktiengesellschaft Synthetic contrast-enhanced mr images
WO2023118044A1 (en) 2021-12-22 2023-06-29 Guerbet Method for processing at least a pre-contrast image and a contrast image respectively depicting a body part prior to and after an injection of a first dose of contrast agent
EP4202855A1 (en) 2021-12-22 2023-06-28 Guerbet Method for processing at least a pre-contrast image and a contrast image respectively depicting a body part prior to and after an injection of a dose of contrast agent
WO2023135056A1 (en) 2022-01-11 2023-07-20 Bayer Aktiengesellschaft Synthetic contrast-enhanced ct images
EP4210069A1 (en) 2022-01-11 2023-07-12 Bayer Aktiengesellschaft Synthetic contrast-enhanced ct images
WO2023161041A1 (en) 2022-02-24 2023-08-31 Bayer Aktiengesellschaft Prediction of representations of an examination area of an examination object after applications of different amounts of a contrast agent
EP4233726A1 (en) 2022-02-24 2023-08-30 Bayer AG Prediction of a representation of an area of an object to be examined after the application of different amounts of a contrast agent
WO2024046833A1 (en) 2022-08-30 2024-03-07 Bayer Aktiengesellschaft Generation of synthetic radiological images
WO2024046831A1 (en) 2022-08-30 2024-03-07 Bayer Aktiengesellschaft Generation of synthetic radiological images
EP4336513A1 (en) 2022-08-30 2024-03-13 Bayer Aktiengesellschaft Generation of synthetic radiological recordings
EP4332601A1 (en) 2022-09-05 2024-03-06 Bayer AG Generation of artificial contrast agent-enhanced radiological recordings
WO2024052156A1 (en) 2022-09-05 2024-03-14 Bayer Aktiengesellschaft Generation of artificial contrast-enhanced radiological images
EP4369353A1 (en) 2022-11-12 2024-05-15 Bayer Aktiengesellschaft Generation of artificial contrast agent-enhanced radiological recordings
EP4369285A1 (en) 2022-11-12 2024-05-15 Bayer AG Generation of artificial contrast agent-enhanced radiological recordings
WO2024100234A1 (en) 2022-11-12 2024-05-16 Bayer Aktiengesellschaft Generation of artificial contrast-enhanced radiological images
WO2024100233A1 (en) 2022-11-12 2024-05-16 Bayer Aktiengesellschaft Generation of artificial contrast-enhanced radiological images

Also Published As

Publication number Publication date
JP2020536638A (en) 2020-12-17
KR20200063222A (en) 2020-06-04
AU2018346938A1 (en) 2020-04-23
CN111601550B (en) 2023-12-05
US20210241458A1 (en) 2021-08-05
US11935231B2 (en) 2024-03-19
US10997716B2 (en) 2021-05-04
EP3694413A4 (en) 2021-06-30
EP3694413A1 (en) 2020-08-19
CA3078728A1 (en) 2019-04-18
AU2018346938B2 (en) 2024-04-04
US20190108634A1 (en) 2019-04-11
JP7244499B2 (en) 2023-03-22
SG11202003232VA (en) 2020-05-28
JP7476382B2 (en) 2024-04-30
BR112020007105A2 (en) 2020-09-24
JP2023078236A (en) 2023-06-06
CN111601550A (en) 2020-08-28

Similar Documents

Publication Publication Date Title
US11935231B2 (en) Contrast dose reduction for medical imaging using deep learning
Hu et al. Bidirectional mapping generative adversarial networks for brain MR to PET synthesis
Zeng et al. Simultaneous single-and multi-contrast super-resolution for brain MRI images based on a convolutional neural network
AU2019234674B2 (en) Systems and methods for generating thin image slices from thick image slices
US20210150671A1 (en) System, method and computer-accessible medium for the reduction of the dosage of gd-based contrast agent in magnetic resonance imaging
KR20230129195A (en) Dose reduction for medical imaging using deep convolutional neural networks
JP2022550688A (en) Systems and methods for improving low-dose volume-enhanced MRI
Cackowski et al. ImUnity: a generalizable VAE-GAN solution for multicenter MR image harmonization
Jun et al. Parallel imaging in time‐of‐flight magnetic resonance angiography using deep multistream convolutional neural networks
Kim et al. Fat-saturated image generation from multi-contrast MRIs using generative adversarial networks with Bloch equation-based autoencoder regularization
Faisal Image Inpainting to Improve the Registration Performance of Multiple Sclerosis (MS) Patient Brain with Brain Atlas
Izadi et al. Enhanced direct joint attenuation and scatter correction of whole-body PET images via context-aware deep networks
CN112700380A (en) PET image volume correction method based on MR gradient information and deep learning
Zhao et al. Using anatomic magnetic resonance image information to enhance visualization and interpretation of functional images: a comparison of methods applied to clinical arterial spin labeling images
Sui et al. Gradient-guided isotropic MRI reconstruction from anisotropic acquisitions
Tohidi et al. Joint synthesis of WMn MPRAGE and parameter maps using deep learning and an imaging equation
Lecoeur et al. Improving white matter tractography by resolving the challenges of edema
Mao et al. PET parametric imaging based on MR frequency-domain texture information
US20230136320A1 (en) System and method for control of motion in medical images using aggregation
Ikuta et al. Super-Resolution for Brain MR Images from a Significantly Small Amount of Training Data. Comput
Ding et al. Medical Image Quality Assessment
CN112150401A (en) Method for enhancing image quality based on deep learning
Bai et al. DD-WGAN: Generative Adversarial Networks with Wasserstein Distance and Dual-Domain Discriminators for Low-Dose CT
Gan et al. Pseudo-MRI-Guided PET Image Reconstruction Method Based on a Diffusion Probabilistic Model
WO2023041748A1 (en) Apparatus and method for generating a perfusion image, and method for training an artificial neural network therefor

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18866524

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 3078728

Country of ref document: CA

Ref document number: 2020520027

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2018346938

Country of ref document: AU

Date of ref document: 20181009

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 20207013155

Country of ref document: KR

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 2018866524

Country of ref document: EP

Effective date: 20200511

REG Reference to national code

Ref country code: BR

Ref legal event code: B01A

Ref document number: 112020007105

Country of ref document: BR

ENP Entry into the national phase

Ref document number: 112020007105

Country of ref document: BR

Kind code of ref document: A2

Effective date: 20200408