Abstract: For three-dimensional segmentation from two-dimensional intracardiac echocardiography imaging, the three-dimension segmentation is output by a machine-learnt multi-task generator. Rather than the brute force approach of training the generator from 2D ICE images to output a 2D segmentation, the generator is trained from 3D information, such as a sparse ICE volume assembled from the 2D ICE images. Where sufficient ground truth data is not available, computed tomography or magnetic resonance data may be used as the ground truth for the sample sparse ICE volumes. The generator is trained to output both the 3D segmentation and a complete volume (i.e., more voxels represented than in the sparse ICE volume). The 3D segmentation may be further used to project to 2D as an input with an ICE image to another network trained to output a 2D segmentation for the ICE image. Display of the 3D segmentation and/or 2D segmentation may guide ablation of tissue in the patient.

Abstract: For three-dimensional segmentation from two-dimensional intracardiac echocardiography imaging, the three-dimension segmentation is output by a machine-learnt multi-task generator. Rather than the brute force approach of training the generator from 2D ICE images to output a 2D segmentation, the generator is trained from 3D information, such as a sparse ICE volume assembled from the 2D ICE images. Where sufficient ground truth data is not available, computed tomography or magnetic resonance data may be used as the ground truth for the sample sparse ICE volumes. The generator is trained to output both the 3D segmentation and a complete volume (i.e., more voxels represented than in the sparse ICE volume). The 3D segmentation may be further used to project to 2D as an input with an ICE image to another network trained to output a 2D segmentation for the ICE image. Display of the 3D segmentation and/or 2D segmentation may guide ablation of tissue in the patient.

Abstract: Systems and methods for smart selection of training data sets by using clinically driven application dependent evaluation metrics to assess the performance of deep learning models after deployment in the field. A machine trained model is deployed to a clinical environment. An evaluation metric is acquired that correlates with a clinical outcome for each instance of the machine trained model performing the task for a medical procedure. Data sets are flagged that are challenging for the machine trained model based on the evaluation metrics. The flagged data sets are prioritized during retraining of the machine trained model.

Abstract: For three-dimensional segmentation from two-dimensional intracardiac echocardiography imaging, the three-dimension segmentation is output by a machine-learnt multi-task generator. Rather than the brute force approach of training the generator from 2D ICE images to output a 2D segmentation, the generator is trained from 3D information, such as a sparse ICE volume assembled from the 2D ICE images. Where sufficient ground truth data is not available, computed tomography or magnetic resonance data may be used as the ground truth for the sample sparse ICE volumes. The generator is trained to output both the 3D segmentation and a complete volume (i.e., more voxels represented than in the sparse ICE volume). The 3D segmentation may be further used to project to 2D as an input with an ICE image to another network trained to output a 2D segmentation for the ICE image. Display of the 3D segmentation and/or 2D segmentation may guide ablation of tissue in the patient.

Abstract: The disclosure relates to a method for determining a boundary about an area of interest in an image set. The includes obtaining the image set from an imaging modality and processing the image set in a convolutional neural network. The convolutional neural network is trained to perform the acts of predicting an inverse distance map for the actual boundary in the image set; and deriving the boundary from the inverse distance map. The disclosure also relates to a method of training a convolutional neural network for use in such a method, and a medical imaging arrangement.

Abstract: Systems and methods for generating an updated segmentation of an initial segmentation are provided. An initial segmentation of an anatomical object from an input medical image is received. User input modifying the initial segmentation is received. An updated segmentation of the anatomical object in the input medical image is generated using a machine learning based network based on at least one of the initial segmentation and the user input. The updated segmentation is output.

Abstract: For three-dimensional segmentation from two-dimensional intracardiac echocardiography imaging, the three-dimension segmentation is output by a machine-learnt multi-task generator. The machine-learnt multi-task generator is trained from 3D information, such as a sparse ICE volume assembled from the 2D ICE images. The machine-learnt multi-task generator is trained to output both the 3D segmentation and a complete volume. The 3D segmentation may be used to project to 2D as an input with an ICE image to another network trained to output a 2D segmentation for the ICE image. Display of the 3D segmentation and/or 2D segmentation may guide ablation of tissue in the patient.

Abstract: A computer-implemented method for updating a boundary segmentation, the method including receiving image data and an original boundary segmentation including a plurality of boundary points. A plurality of edges in the image data is detected and used to generate an edge map. A confidence for a boundary point in the original boundary segmentation is computed where the confidence is based on a distance between the boundary point and an edge point associated with at least one of the plurality of edges of the edge map, and based on the confidence a classification of the boundary point is determined. An updated boundary segmentation based on the classification of the boundary point is generated and then output.

Abstract: The disclosure relates to a method for determining a boundary about an area of interest in an image set. The includes obtaining the image set from an imaging modality and processing the image set in a convolutional neural network. The convolutional neural network is trained to perform the acts of predicting an inverse distance map for the actual boundary in the image set; and deriving the boundary from the inverse distance map. The disclosure also relates to a method of training a convolutional neural network for use in such a method, and a medical imaging arrangement.

Abstract: For three-dimensional segmentation from two-dimensional intracardiac echocardiography imaging, the three-dimension segmentation is output by a machine-learnt multi-task generator. Rather than the brute force approach of training the generator from 2D ICE images to output a 2D segmentation, the generator is trained from 3D information, such as a sparse ICE volume assembled from the 2D ICE images. Where sufficient ground truth data is not available, computed tomography or magnetic resonance data may be used as the ground truth for the sample sparse ICE volumes. The generator is trained to output both the 3D segmentation and a complete volume (i.e., more voxels represented than in the sparse ICE volume). The 3D segmentation may be further used to project to 2D as an input with an ICE image to another network trained to output a 2D segmentation for the ICE image. Display of the 3D segmentation and/or 2D segmentation may guide ablation of tissue in the patient.

Abstract: A method and apparatus for extracting centerline representations of vascular structures in medical images is disclosed. A vessel orientation tensor for each of a plurality of voxels associated with the target vessel, such as a coronary artery, in a medical image, such as a coronary tomography angiography (CTA) image, using a trained vessel orientation tensor classifier. A flow field is estimated for the plurality of voxels associated with the target vessel in the medical image based on the vessel orientation tensor estimated for each of the plurality of voxels. A centerline of the target vessel is extracted based on the estimated flow field for the plurality of vessels associated with the target vessel in the medical image by detecting a path that carries maximum flow.

Abstract: A method and apparatus for machine learning based detection of vessel orientation tensors of a target vessel from a medical image is disclosed. For each of a plurality of voxels in a medical image, such as a computed tomography angiography (CTA), features are extracted from sampling patches oriented to each of a plurality of discrete orientations in the medical image. A classification score is calculated for each of the plurality of discrete orientations at each voxel based on the features extracted from the sampling patches oriented to each of the plurality of discrete orientations using a trained vessel orientation tensor classifier. A vessel orientation tensor at each voxel is calculated based on the classification scores of the plurality of discrete orientations at that voxel.

Abstract: A method and apparatus for learning based classification of vascular branches to distinguish falsely detected branches from true branches is disclosed. A plurality of overlapping fixed size branch segments are sampled from branches of a detected centerline tree of a target vessel extracted from a medical image of a patient. A plurality of 1D profiles are extracted along each of the overlapping fixed size branch segments. A probability score for each of the overlapping fixed size branch segments is calculated based on the plurality of 1D profiles extracted for each branch segment using a trained deep neural network classifier. The trained deep neural network classifier may be a convolutional neural network (CNN) trained to predict a probability of a branch segment being fully part of a target vessel based on multi-channel 1D input.

Abstract: A method and apparatus for learning based classification of vascular branches to distinguish falsely detected branches from true branches is disclosed. A plurality of overlapping fixed size branch segments are sampled from branches of a detected centerline tree of a target vessel extracted from a medical image of a patient. A plurality of 1D profiles are extracted along each of the overlapping fixed size branch segments. A probability score for each of the overlapping fixed size branch segments is calculated based on the plurality of 1D profiles extracted for each branch segment using a trained deep neural network classifier. The trained deep neural network classifier may be a convolutional neural network (CNN) trained to predict a probability of a branch segment being fully part of a target vessel based on multi-channel 1D input.

Abstract: A method and apparatus for machine learning based detection of vessel orientation tensors of a target vessel from a medical image is disclosed. For each of a plurality of voxels in a medical image, such as a computed tomography angiography (CTA), features are extracted from sampling patches oriented to each of a plurality of discrete orientations in the medical image. A classification score is calculated for each of the plurality of discrete orientations at each voxel based on the features extracted from the sampling patches oriented to each of the plurality of discrete orientations using a trained vessel orientation tensor classifier. A vessel orientation tensor at each voxel is calculated based on the classification scores of the plurality of discrete orientations at that voxel.

Abstract: A method and apparatus for extracting centerline representations of vascular structures in medical images is disclosed. A vessel orientation tensor for each of a plurality of voxels associated with the target vessel, such as a coronary artery, in a medical image, such as a coronary tomography angiography (CTA) image, using a trained vessel orientation tensor classifier. A flow field is estimated for the plurality of voxels associated with the target vessel in the medical image based on the vessel orientation tensor estimated for each of the plurality of voxels. A centerline of the target vessel is extracted based on the estimated flow field for the plurality of vessels associated with the target vessel in the medical image by detecting a path that carries maximum flow.

Abstract: The left ventricle epicardium is estimated in medical diagnostic imaging. C-arm x-ray data is used to detect an endocardium at different phases. The detected endocardium at the different phases is compared to sample endocardiums at different phases. The sample endocardiums have corresponding sample epicardiums. The transformation between the most similar sample endocardium or endocardiums over time and the detected endocardium over time is applied to the corresponding sample epicardium or epicardiums. The transformed sample epicardium over time is the estimated epicardium over time for the C-arm x-ray data.

Abstract: A method and system for contrast inflow detection in a sequence of fluoroscopic images is disclosed. Vessel segments are detected in each frame of a fluoroscopic image sequence. A score vector is determined for the fluoroscopic image sequence based on the detected vessel segments in each frame of the fluoroscopic image sequence. It is determined whether a contrast agent injection is present in the fluoroscopic image sequence based on the score vector. If it is determined that a contrast agent injection is present in the fluoroscopic image sequence, a contrast inflow frame, at which contrast agent inflow begins, is detected in the fluoroscopic image sequence based on the score vector.

Abstract: A system and method for regression-based segmentation of the mitral valve in 2D+t cardiac magnetic resonance (CMR) slices is disclosed. The 2D+t CMR slices are acquired according to a mitral valve-specific acquisition protocol introduced herein. A set of mitral valve landmarks is detected in each 2D CMR slice and mitral valve contours are estimated in each 2D CMR slice based on the detected landmarks. A full mitral valve model is reconstructed from the mitral valve contours estimated in the 2D CMR slices using a trained regression model. Each 2D CMR slice may be a cine image acquired over a full cardiac cycle. In this case, the segmentation method reconstructs a patient-specific 4D dynamic mitral valve model from the 2D+t CMR image data.

Abstract: Background information is subtracted from projection data in medical diagnostic imaging. The background is removed using data acquired in a single rotational sweep of a C-arm. The removal may be by masking out a target, leaving the background, in the data as constructed into a volume. For subtraction, the masked background information is projected to a plane and subtracted from the data representing the plane.