Abstract: A process can include performing meta-learning for a variance-aware prototypical network pre-trained on a dataset comprising examples of a first type of radiology report associated with a single domain. The meta-learning comprises learning one or more prototype representations for each radiology classification task and a variance information for the prototype representations of each radiology classification task. The one or more respective prototype representations for each radiology classification task are modeled as a Gaussian and a query sample comprising text data of a type of radiology report seen during the meta-learning is provided to the variance-aware prototypical network. A distance metric is determined between a Dirac distribution representation of the query sample and the Gaussians of the respective prototype representations for each radiology classification task included in the meta-learning.

Abstract: Described are systems, techniques, and processes for pathology detection in radiological images. A process can include obtaining a radiological image corresponding to an imaged anatomical area. Based on processing the radiological image using a semantic segmentation neural network, a target map can be generated corresponding to a plurality of candidate anatomical defect locations within the cropped radiological image. At least one volume of interest (VOI) can be generated centered around a particular candidate anatomical defect location within the cropped radiological image. A classification neural network can be used to classify the particular candidate anatomical defect location within the cropped radiological image, wherein classifying the particular candidate anatomical defect location includes determining a pathology associated with the particular candidate anatomical defect location.

Abstract: Described are techniques for image processing. For instance, a process can include obtaining a plurality of labeled input images and determining a threshold percentage associated with the plurality of labeled input images, indicative of a percentage of correctly labeled input images. The process can include determining a respective self-influence for each respective labeled input image included in the plurality of input images and generating a respective self-influence weight for each respective labeled input image, based on the respective self-influence and the threshold percentage associated with each respective labeled input image. The process can include determining one or more loss values using a loss function associated with training a machine learning network based on using the plurality of labeled input images as a training data set, wherein the loss function determines the one or more loss values based on weighting each respective labeled input image by its respective self-influence weight.

Abstract: Described are techniques for uncertainty-aware anatomical landmark detection, using, for example, a deep reinforcement learning (DRL) anatomical landmark detection agent. For instance, a process can include generating one or more image features for an input medical image using a first sub-network of the anatomical landmark detection agent. A softmax layer of a second sub-network of the anatomical landmark detection agent can generate a plurality of discrete Q-value distributions for a set of allowable actions associated with movement of the agent within the medical image. An anatomical landmark location within the medical image can be predicted using the discrete Q-value distributions. An uncertainty can be determined for the predicted anatomical landmark location, based on an average full width half maximum (FWHM) calculated for the plurality of discrete Q-value distributions.

Abstract: A method for visualizing airway wall abnormalities includes acquiring\Dual Energy Computed Tomography (DECT) imaging data comprising one or more image volumes representative of a bronchial tree. An iodine map is derived using the DECT imaging data and the bronchial tree is segmented from the image volume(s). A tree model representative of the bronchial tree is generated. Then, for each branch, this tree model is used to determine an indicator of normal or abnormal thickness. Locations corresponding to bronchial walls in the bronchial tree using the tree model are identified. Next, for each branch, the locations corresponding to bronchial walls in the bronchial tree and the iodine map are used to determine an indicator of normal or abnormal inflammation. A visualization of the bronchial tree may be presented with visual indicators at each of the locations corresponding to bronchial walls indicating whether a bronchial wall is thickened and/or inflamed.

Abstract: A method for visualizing airway wall abnormalities includes acquiring \Dual Energy Computed Tomography (DECT) imaging data comprising one or more image volumes representative of a bronchial tree. An iodine map is derived using the DECT imaging data and the bronchial tree is segmented from the image volume(s). A tree model representative of the bronchial tree is generated. Then, for each branch, this tree model is used to determine an indicator of normal or abnormal thickness. Locations corresponding to bronchial walls in the bronchial tree using the tree model are identified. Next, for each branch, the locations corresponding to bronchial walls in the bronchial tree and the iodine map are used to determine an indicator of normal or abnormal inflammation. A visualization of the bronchial tree may be presented with visual indicators at each of the locations corresponding to bronchial walls indicating whether a bronchial wall is thickened and/or inflamed.

Abstract: Traumatic brain injury (TBI) in a patient is assessed. A TBI diagnosis for the patient is determined based on features from MRI data, such as anatomical features, functional features, diffusion features, connectivity features from functional MRI, connectivity features from diffusion MRI, and/or network features from the connectivity features. The TBI diagnosis is determined using a trained classifier. The classifier synthesizes the features into a single number (e.g., a confidence in the prediction of the diagnosis) and indicates the features most responsible for the diagnosis. The disease trajectory for a given patient may be predicted using the trained classifier.

Abstract: Traumatic brain injury (TBI) in a patient is assessed. A TBI diagnosis for the patient is determined based on features from MRI data, such as anatomical features, functional features, diffusion features, connectivity features from functional MRI, connectivity features from diffusion MRI, and/or network features from the connectivity features. The TBI diagnosis is determined using a trained classifier. The classifier synthesizes the features into a single number (e.g., a confidence in the prediction of the diagnosis) and indicates the features most responsible for the diagnosis. The disease trajectory for a given patient may be predicted using the trained classifier.

Abstract: A method for grouping airway and artery pairs, includes: computing a two-dimensional (2D) cross-section of an airway; identifying regions of high-intensity in the 2D cross-section; computing a first indicator for each of the high intensity regions, wherein the first indicator is an orientation measure of the high intensity region with respect to the airway; computing a second indicator for each of the high intensity regions, wherein the second indicator is a circularity measure of the high intensity region; computing a third indicator for each of the high intensity regions, wherein the third indicator is a proximity measure of the high intensity region with respect to the airway; summing the first through third indicators for each of the high intensity regions to obtain a score for each of the high intensity regions; and determining which of the high intensity regions is an artery corresponding to the airway based on its score.

Abstract: A method for visualizing airways in chest images, includes: computing a distance map of a segmented bronchial tree; extracting data from the segmented bronchial tree using the distance map; and visualizing a three-dimensional (3D) image of the segmented bronchial tree color-coded according to the extracted data.

Abstract: A branch extension method and system for segmenting airways in 3D image data is disclosed. An initial airway segmentation is obtained from the 3D image data. Terminal branches of segmented airways of the initial airway segmentation are identified. The segmentation of the terminal branches is then extended. The segmentation of the terminal branches can be extended using various segmentation techniques. This method can use complex segmentation techniques to extend the terminal branches without having a large impact to the overall speed of the segmentation.

Abstract: A method for determining a size of an airway lumen and a thickness of an airway wall includes: computing a centerline of an airway; computing a three-dimensional (3D) gradient of a volume of the airway within a first threshold; positioning a tube along the centerline; iteratively expanding the tube by increasing its radius until the radius of the tube reaches the first threshold; determining inner and outer radii of the tube by checking the 3D gradient computed along an x-axis and a y-axis of the tube at a boundary of the tube at each iteration; and fitting the tube to the airway by using the determined inner and outer radii, wherein the inner radius of the fit tube is half a diameter of the airway lumen and the outer radius of the fit tube minus the inner radius of the fit tube is a thickness of the airway wall.

Abstract: A method for segmenting a solid component (SC) in a ground glass nodule (GGN) includes providing a digitized image that includes a segmented GGN, the image comprising a plurality of intensities corresponding to a 3-dimensional grid of points, computing an intensity threshold that distinguishes a high intensity solid component of the GGN from a low intensity non-solid component, and applying the intensity threshold to identify a SC of the GGN and to identify regions of interest around the GGN, detecting whether or not a region of interest is a vessel, calculating a model for a detected vessel based on a radius and orientation of the vessel, and removing from the GGN segmentation all points that belong to both the model and the SC inside the GGN, and verifying whether a structure resulting from excluding the points qualifies as an SC.

Abstract: A method for detecting and localizing mucus plugs in digitized lung images, includes providing a digitized lung image volume comprising a plurality of intensities corresponding to a 3-dimensional grid of points, extracting a bronchial tree from said lung image, said bronchial tree comprising a plurality of branching airways terminating at terminal points, providing a model of a 2-dimensional cross section of an airway, selecting an extended point beyond a terminal point of an airway branch in a direction of said airway branch, obtaining a 2-dimensional cross section I of size m×n points from said lung image about said selected point, processing said 2D cross section I by calculating a local neighborhood function for each point in the cross section and forming a union of all local neighborhood functions, and calculating a correlation between processed 2D cross section and said airway model, wherein said correlation is indicative of the presence of a mucus plug within said airway.

Abstract: A method and system for detecting airways in 3D lung image data is disclosed. The 3D lung image data is filtered using one or more filters based on first and second order derivatives of the CT image data. Each filter calculates a value for each voxel of the 3D lung image data, and the values from all of the filters are combined to determine a voxel score for each voxel. If the voxel score for a voxel is greater than or equal to a threshold value, the voxel is considered an airway candidate.

Abstract: A computer-implemented method for intervertebral disc alignment includes providing a spine image and a click point on the spine image, applying an adaptive thresholding technique to segment vertebrae regions at the click point from the spine image, and filtering segmented vertebrae regions with a morphological operation. The method further includes estimating a multi-scale orientation field from filtered segmented vertebrae regions, extracting an intervertebral disc region by applying region growing for each scale, integrating the multi-scale orientation field in a locally segmented intervertebral disc region based on the intervertebral disc region, estimating a disc centerline from the filtered segmented vertebrae regions, and fusing an integrated multi-scale orientation field with the disc centerline based on associated confidence measures to provide an intervertebral disc alignment.

Abstract: Disclosed is a systematic way of automatically segmenting lung regions. To increase the efficiency of a lung segmentation technique, a region-based technique, such as region growing, is performed by a computer on a middle slice of the CT volume. A contour-based technique is then used for a plurality of non-middle slices of the CT volume. This allows the implementation to be multithreaded and results in an improvement in the segmentation algorithm's efficiency.

Abstract: A method for providing automatic detection of curvature of a spine and computation of specific angles in images of the spine includes automatically displaying the curvature of the spine as a line in an image of the spine, and computing at least one of a first angle or a second angle based on the line of the curvature of the spine.

Abstract: A computer-implemented method for vertebrae segmentation includes providing an image of a plurality of vertebrae, and determining a seed in each of at least two adjacent vertebrae in the image. The method further includes mapping a unit square to the seeds in the image as corresponding shape constraints on a segmentation, evolving the shape constraints to determine the segmentation of the adjacent vertebrae, wherein evolutions of the shape constraints interact, and outputting a segmented image indicating a location of the vertebra.

Abstract: We present an algorithm for local surface smoothing in a defined Volume of Interest (“VOI”) cropped from three-dimensional (“3D”) volume data, such as lung computer tomography (“CT”) data. Because the VOI is generally a smooth and piecewise linear surface, the inclusion of one or more bumps may suggest an abnormality. In lung CT data, for example, such bumps can be nodules that are grown from the chest wall. The nodules may represent a possibility of lung cancer. Through surface smoothing, potential pathologies are separated from the surrounding anatomical structures. For example, nodules may be segmented from the chest wall. The separated pathologies can be analyzed as diagnostic evidence.