Abstract: Methods and apparatus for automated medical image analysis using deep learning networks are disclosed. In a method of automatically performing a medical image analysis task on a medical image of a patient, a medical image of a patient is received. The medical image is input to a trained deep neural network. An output model that provides a result of a target medical image analysis task on the input medical image is automatically estimated using the trained deep neural network. The trained deep neural network is trained in one of a discriminative adversarial network or a deep image-to-image dual inverse network.

Abstract: Intelligent multi-scale image parsing determines the optimal size of each observation by an artificial agent at a given point in time while searching for the anatomical landmark. The artificial agent begins searching image data with a coarse field-of-view and iteratively decreases the field-of-view to locate the anatomical landmark. After searching at a coarse field-of view, the artificial agent increases resolution to a finer field-of-view to analyze context and appearance factors to converge on the anatomical landmark. The artificial agent determines applicable context and appearance factors at each effective scale.

Abstract: Methods and apparatus for automated medical image analysis using deep learning networks are disclosed. In a method of automatically performing a medical image analysis task on a medical image of a patient, a medical image of a patient is received. The medical image is input to a trained deep neural network. An output model that provides a result of a target medical image analysis task on the input medical image is automatically estimated using the trained deep neural network. The trained deep neural network is trained in one of a discriminative adversarial network or a deep image-to-image dual inverse network.

Abstract: A learning-based magnetic resonance fingerprinting (MRF) reconstruction method for reconstructing an MR image of a tissue space in an MR scan subject for a particular MR sequence is disclosed. The method involves using a machine-learning algorithm that has been trained to generate a set of tissue parameters from acquired MR signal evolution without using a dictionary or dictionary matching.

Abstract: Embodiments can provide a method for atlas-based contouring, comprising constructing a relevant atlas database; selecting one or more optimal atlases from the relevant atlas database; propagating one or more atlases; fusing the one or more atlases; and assessing the quality of one or more propagated contours.

Abstract: A method and apparatus for automated liver segmentation in a 3D medical image of a patient is disclosed. A 3D medical image, such as a 3D computed tomography (CT) volume, of a patient is received. The 3D medical image of the patient is input to a trained deep image-to-image network. The trained deep image-to-image network is trained in an adversarial network together with a discriminative network that distinguishes between predicted liver segmentation masks generated by the deep image-to-image network from input training volumes and ground truth liver segmentation masks. A liver segmentation mask defining a segmented liver region in the 3D medical image of the patient is generated using the trained deep image-to-image network.

Abstract: Systems and methods are provided for automatic detection and quantification for traumatic bleeding. Image data is acquired using a full body dual energy CT scanner. A machine-learned network detects one or more bleeding areas on a bleeding map from the dual energy CT scan image data. A visualization is generated from the bleeding map. The predicted bleeding areas are quantified, and a risk value is generated. The visualization and risk value are presented to an operator.

Abstract: A method and apparatus for deep learning based automatic bone removal in medical images, such as computed tomography angiography (CTA) volumes, is disclosed. Bone structures are segmented in a 3D medical image of a patient by classifying voxels of the 3D medical image as bone or non-bone voxels using a deep neural network trained for bone segmentation. A 3D visualization of non-bone structures in the 3D medical image is generated by removing voxels classified as bone voxels from a 3D visualization of the 3D medical image.

Abstract: A computer-implemented method for identifying features in 3D image volumes includes dividing a 3D volume into a plurality of 2D slices and applying a pre-trained 2D multi-channel global convolutional network (MC-GCN) to the plurality of 2D slices until convergence. Following convergence of the 2D MC-GCN, a plurality of parameters are extracted from a first feature encoder network in the 2D MC-GCN. The plurality of parameters are transferred to a second feature encoder network in a 3D Anisotropic Hybrid Network (AH-Net). The 3D AH-Net is applied to the 3D volume to yield a probability map;. Then, using the probability map, one or more of (a) coordinates of the objects with non-maximum suppression or (b) a label map of objects of interest in the 3D volume are generated.

Abstract: A method of reviewing neural scans includes receiving at least one landmark corresponding to an anatomical region. A plurality of images of tissue including the anatomical region is received and a neural network configured to differentiate between healthy tissue and unhealthy tissue within the anatomical region is generated. The neural network is generated by a machine learning process configured to receive the plurality of images of tissue and generate a plurality of weighting factors configured to differentiate between healthy tissue and unhealthy tissue. At least one patient image of tissue including the anatomical region is received and a determination is made by the neural network whether the at least one patient image of tissue includes healthy or unhealthy tissue.

Abstract: A method and apparatus for detecting vascular landmarks in a 3D image volume, such as a CT volume, is disclosed. One or more guide slices are detected in a 3D image volume. A set of landmark candidates for multiple target vascular landmarks are then detected based on the guide slices. A node potential value for each landmark candidate is generated based on an error value determined using spatial histogram-based error regression, and edge potential values for pairs of landmark candidates are generated based on a bifurcation analysis of the image volume using vessel tracing. The optimal landmark candidate for each target landmark is then determined using a Markov random field model based on the node potential values and the edge potential values.

Abstract: Methods and apparatus for cross-domain medical image analysis and cross-domain medical image synthesis using deep image-to-image networks and adversarial networks are disclosed. In a method for cross-domain medical image analysis a medical image of a patient from a first domain is received. The medical image is input to a first encoder of a cross-domain deep image-to-image network (DI2IN) that includes the first encoder for the first domain, a second encoder for a second domain, and a decoder. The first encoder converts the medical image to a feature map and the decoder generates an output image that provides a result of a medical image analysis task from the feature map.

Abstract: A medical imaging phantom (18) is three-dimensionally printed (36). In one specific approach, three-dimensional printing (36) allows for any number of variations in phantoms (18). A library of different phantoms (18), different inserts, different textures, different densities, different organs, different pathologies, different sizes, different shapes, and/or other differences allows for defining a specific phantom (18) as needed. The defined phantom (18) is then printed (36) for calibration or other use in medical imaging.

Abstract: A method of visualizing spinal nerves includes receiving a 3D image volume depicting a spinal cord and a plurality of spinal nerves. For each spinal nerve, a 2D spinal nerve image is generated by defining a surface within the 3D volume comprising the spinal nerve. The surface is curved such that it passes through the spinal cord while encompassing the spinal nerve. Then, the 2D spinal nerve images are generated based on voxels on the surface included in the 3D volume. A visualization of the 2D spinal images is presented in a graphical user interface that allows each 2D spinal image to be viewed simultaneously.

Abstract: A computer-implemented method for generating contours of anatomy based on user click points includes a computer displaying an image comprising an anatomical structure and receiving a first user selection of a first click point at a first position on an outward facing edge of the anatomical structure. The computer applies a contour inference algorithm to generate an inferred contour around the outward facing edge based on the first position. Following generation of the inferred contour, the computer receives a second user selection of a second click point at a second position on the image. Then, the computer creates a visual indicator on a segment of the inferred contour between the first position and the second position as indicative of the user's confirmation of accuracy of the segment.

Abstract: A method and system for anatomical landmark detection in medical images using deep neural networks is disclosed. For each of a plurality of image patches centered at a respective one of a plurality of voxels in the medical image, a subset of voxels within the image patch is input to a trained deep neural network based on a predetermined sampling pattern. A location of a target landmark in the medical image is detected using the trained deep neural network based on the subset of voxels input to the trained deep neural network from each of the plurality of image patches.

Abstract: In scan data retrieval, a mesh is fit (32) to surface data of a current patient, such as data from an optical or depth sensor (18). Meshes are also fit (48) to medical scan data, such as fitting (48) to skin surface segments of computed tomography data. The meshes or parameters derived from the meshes may be more efficiently compared (34) to identify (36) a previous patient with similar body shape and/or size. The scan configuration (38) for that patient, or that patient as altered to account for differences from the current patient, is used. In some embodiments, the parameter vector used for searching (34) includes principle component analysis coefficients. In further embodiments, the principle component analysis coefficients may be projected to a more discriminative space using metric learning.

Abstract: A method of classifying signals using non-linear sparse representations includes learning a plurality of non-linear dictionaries based on a plurality of training signals, each respective nonlinear dictionary corresponding to one of a plurality of class labels. A non-linear sparse coding process is performed on a test signal for each of the plurality of non-linear dictionaries, thereby associating each of the plurality of non-linear dictionaries with a distinct sparse coding of the test signal. For each respective non-linear dictionary included in the plurality of non-linear dictionaries, a reconstruction error is measured using the test signal and the distinct sparse coding corresponding to the respective non-linear dictionary. A particular nonlinear dictionary corresponding to a smallest value for the reconstruction error among the plurality of non-linear dictionaries is identified and a class label corresponding to the particular non-linear dictionary is assigned to the test signal.

Abstract: The process of creating a 3D printer ready model of patient specific anatomy is automated. Instead of manual manipulation of the 3D mesh from imaging to create the 3D printer ready model, an automated manipulation is provided. The segmentation may be automated as well. In one approach, a transform between a predetermined mesh of anatomy and patient specific 3D mesh is calculated. The predetermined mesh has a corresponding 3D printer ready model. By applying the transform to the 3D printer ready model, the 3D printer ready model is altered to become specific to the patient. In addition, target manipulation that alters semantic parts of the anatomical structure may be included in 3D printing.

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.