Abstract: Described herein are systems, methods, and instrumentalities associated with image segmentation such as tubular structure segmentation. An artificial neural network is trained to segment tubular structures of interest in a medical scan image based on annotated images of a different type of tubular structures that may have a different contrast and/or appearance from the tubular structures of interest. The training may be conducted in multiple stages during which a segmentation model learned from the annotated images during a first stage may be modified to fit the tubular structures of interest in a second stage. In examples, the tubular structures of interest may include coronary arteries, catheters, guide wires, etc., and the annotated images used for training the artificial neural network may include blood vessels such as retina blood vessels.

Abstract: Described herein are systems, methods, and instrumentalities associated with denoising medical images such as fluoroscopic images using deep learning techniques. A first artificial neural network (ANN) is trained to denoise an input medical image in accordance with a provided target noise level. The training of the first ANN is conducted by pairing a noisy input image with target denoised images that include different levels of noise. These target denoised images are generated using a second ANN as intermediate outputs of the second ANN during different training iterations. As such, the first ANN may learn to perform the denoising task in an unsupervised manner without requiring noise-free training images as the ground truth.

Abstract: Described herein are systems, methods, and instrumentalities associated with landmark detection. The detection may be accomplished by determining a graph representation of a plurality of hypothetical landmarks detected in one or more medical images. The graph representation may include nodes that represent the hypothetical landmarks and edges that represent the relationships between paired hypothetical landmarks. The graph representation may be processed using a graph neural network such a message passing graph neural network, by which the landmark detection problem may be converted and solved as a graph node labeling problem.

Abstract: Described herein are systems, methods and instrumentalities associated with automatic assessment of aneurysms. An automatic aneurysm assessment system or apparatus may be configured to obtain, e.g., using a pre-trained artificial neural network, strain values associated one or more locations of a human heart and one or more cardiac phases of the human heart and derive a representation (e.g., a 2D matrix) of the strain values across time and/or space. The system or apparatus may determine, based on the derived representation of the strain values, respective strain patterns associated with the one or more locations of the human heart and further determine whether the one or more locations are aneurysm locations by comparing the automatically determined strain patterns with predetermined normal strain patterns of the heart and determining the presence or risk of aneurysms based on the comparison.

Abstract: Described herein are systems, methods, and instrumentalities associated with generating multi-contrast MRI images associated with an MRI study. The systems, methods, and instrumentalities utilize an artificial neural network (ANN) trained to jointly determine MRI data sampling patterns for the multiple contrasts based on predetermined quality criteria associated with the MRI study and reconstruct MRI images with the multiple contrasts based on under-sampled MRI data acquired using the sampling patterns. The training of the ANN may be conducted with an objective to improve the quality of the whole MRI study rather than individual contrasts. As such, the ANN may learn to allocate resources among the multiple contrasts in a manner that optimizes the performance of the whole MRI study.

Abstract: Described herein are systems, methods, and instrumentalities associated with segmenting and/or determining the shape of an anatomical structure. An artificial neural network (ANN) is used to perform these tasks based on a statistical shape model of the anatomical structure. The ANN is trained by evaluating and backpropagating multiple losses associated with shape estimation and segmentation mask generation. The model obtained using these techniques may be used for different clinical purposes including, for example, motion estimation and motion tracking.

Abstract: Described herein are neural network-based systems, methods and instrumentalities associated with cardiac assessment. An apparatus as described herein may obtain electrocardiographic imaging (ECGI) information associated with a human heart and magnetic resonance imaging (MRI) information associated with the human heart, and integrate the ECGI and MRI information using a machine-learned model. Using the integrated ECGI and MRI information, the apparatus may predict target ablation sites, estimate electrophysiology (EP) measurements, and/or simulate the electrical system of the human heart.

Abstract: In Multiplex MRI image reconstruction, a hardware processor acquires sub-sampled Multiplex MRI data and reconstructs parametric images from the sub-sampled Multiplex MRI data. A machine learning model or deep learning model uses the subsampled Multiplex MRI data as the input and parametric maps calculated from the fully sampled data, or reconstructed fully sample data, as the ground truth. The model learns to reconstruct the parametric maps directly from the subsampled Multiplex MRI data.

Abstract: A neural network system implements a model for generating an output image based on a received input image. The model is learned through a training process during which parameters associated with the model are adjusted so as to maximize a difference between a first image predicted using first parameter values of the model and a second image predicted using second parameter values of the model, and to minimize a difference between the second image and a ground truth image. During a first iteration of the training process the first image is predicted and during a second iteration the second image is predicted. The first parameter values are obtained during the first iteration by minimizing a difference between the first image and the ground truth image, and the second parameter values are obtained during the second iteration by maximizing the difference between the first image and the second image.

Abstract: Disclosed herein are systems, methods, and instrumentalities associated with reconstructing magnetic resonance (MR) images based on multi-slice, under-sampled MRI data (e.g., k-space data). The multi-slice MRI data may be acquired using a simultaneous multi-slice (SMS) technique and MRI information associated with multiple MRI slices may be entangled in the multi-slice MRI data. A neural network may be trained and used to disentangle the MRI information and reconstruct MRI images for the different slices. A data consistency component may be used to estimate k-space data based on estimates made by the neural network, from which respective MRI images associated with multiple MRI slices may be obtained by applying a Fourier transform to the k-space data.

Abstract: Deep learning based systems, methods, and instrumentalities are described herein for registering images from a same imaging modality and different imaging modalities. Transformation parameters associated with the image registration task are determined using a neural ordinary differential equation (ODE) network that comprises multiple layers, each configured to determine a respective gradient update for the transformation parameters based on a current state of the transformation parameters received by the layer. The gradient updates determined by the multiple ODE layers are then integrated and applied to initial values of the transformation parameters to obtain final parameters for completing the image registration task. The operations of the ODE network may be facilitated by a feature extraction network pre-trained to determine content features shared by the input images.

Abstract: A system comprising a first computing apparatus in communication with multiple second computing apparatuses. The first computing apparatus may obtain a plurality of first trained machine learning models for a task from the multiple second computing apparatuses. At least a portion of parameter values of the plurality of first trained machine learning models may be different from each other. The first computing apparatus may also obtain a plurality of training samples. The first computing apparatus may further determine, based on the plurality of training samples, a second trained machine learning model by learning from the plurality of first trained machine learning models.

Abstract: Embodiments of the disclosure provide systems and methods for segmenting an image. An exemplary system includes a communication interface configured to receive the image acquired by an image acquisition device. The system further includes a memory configured to store a multi-level learning network comprising a plurality of convolution blocks cascaded at multiple levels. The system also includes a processor configured to apply a first convolution block and a second convolution block of the multi-level learning network to the image in series. The first convolution block is applied to the image and the second convolution block is applied to a first output of the first convolution block. The processor is further configured to concatenate the first output of the first convolution block and a second output of the second convolution block to obtain a feature map and obtain a segmented image based on the feature map.

Abstract: Embodiments of the disclosure provide systems and methods for generating a report based on a medical image of a patient. An exemplary system includes a communication interface configured to receive the medical image acquired by an image acquisition device. The system may further include at least one processor. The at least one processor is configured to automatically determine keywords from a natural language description of the medical image generated by applying a learning network to the medical image. The at least one processor is further configured to generate the report describing the medical image of the patient based on the keywords. The at least one processor is also configured to provide the report for display.

Abstract: Disclosed herein are systems, methods, and instrumentalities associated with reconstructing magnetic resonance (MR) images based on under-sampled MR data. The MR data include 2D or 3D information, and may encompass multiple contrasts and multiple coils. The MR images are reconstructed using deep learning (DL) methods, which may accelerate the scan and/or image generation process. Challenges imposed by the large quantity of the MR data and hardware limitations are overcome by separately reconstructing MR images based on respective subsets of contrasts, coils, and/or readout segments, and then combining the reconstructed MR images to obtain desired multi-contrast results.

Abstract: Neural network based systems, methods, and instrumentalities may be used to remove motion artifacts from magnetic resonance (MR) images. Such a neural network based system may be trained to perform the motion artifact removal tasks without reference (e.g., without using paired motion-contaminated and motion-free MR images). Various training techniques are described herein including one that feeds the neural network with pairs of MR images with different levels of motion contamination and forces the neural network learn to correct the motion contamination by transforming a first image of a contaminated pair into a second image of the contaminated pair. Other neural network training techniques are also described with an aim to reduce the reliance on training data that is difficult to obtain.

Abstract: Methods and systems for classifying an image.

Abstract: Motion contaminated magnetic resonance (MR) images for training an artificial neural network to remove motion artifacts from the MR images are difficult to obtain. Described herein are systems, methods, and instrumentalities for injecting motion artifacts into clean MR images and using the artificially contaminated images for machine learning and neural network training. The motion contaminated MR images may be created based on clean source MR images that are associated with multiple physiological cycles of a scanned object, and by deriving MR data segments for the multiple physiological cycles based on the source MR images. The MR data segments thus derived may be combined to obtain a simulated MR data set, from which one or more target MR images may be generated to exhibit a motion artifact. The motion artifact may be created by manipulating the source MR images and/or controlling the manner in which the MR data set or the target MR images are generated.

Abstract: A bullseyes plot may be generated based on cardiac magnetic resonance imaging (CMRI) to facilitate the diagnosis and treatment of heart diseases. Described herein are systems, methods, and instrumentalities associated with bullseyes plot generation. A plurality of myocardial segments may be obtained for constructing the bullseye plot based on landmark points detected in short-axis and long-axis magnetic resonance (MR) slices of the heart and by arranging the short-axis MR slices sequentially in accordance with the order in which the slices are generated during the CMRI. The sequential order of the short-axis MR slices may be determined utilizing projected locations of the short-axis MR slices on a long-axis MR slice and respective distances of the projected locations to a landmark point of the long-axis MR slice. The myocardium and/or landmark points may be identified in the short-axis and/or long-axis MR slices using artificial neural networks.

Abstract: Described herein are neural network-based systems, methods and instrumentalities associated with estimating a thickness of an anatomical structure based on a visual representation of the anatomical structure and a machine-learned thickness prediction model. The visual representation may include an image or a segmentation mask of the anatomical structure. The thickness prediction model may be learned based on ground truth information derived by applying a partial differential equation such as Laplace's equation to the visual representation and solving the partial differential equation. When the visual representation includes an image of the anatomical structure, the systems, methods and instrumentalities described herein may also be capable of generating a segmentation mask of the anatomical structure based on the image.