National Institutes of Health, a Component of the US Dept. of Health & Human Services

Mayo Foundation for Medical Education and Research

Ronald M Summers

Marek Franaszek

Gheorghe Iordanescu

Scott Selbie

James D Malley

Lynne M Pusanik

Le Lu

Peter Choyke

Jianhua Yao

Baris Turkbey

Bradford Johns Wood

Adam Patrick Harrison

Sheng Xu

Nathan S Lay

C Daniel Johnson

Xiaosong Wang

Peter A Pinto

Ziyue Xu

Matthew Greer

Jin Tae Kwak

Ruida Cheng

Holger Roth

Robert L Van Uitert, Jr

Ingmar Bitter

Matthew J McAuliffe

Jiang Li

Sharon Greenblum

Sonia Gaur

Juan Raul Cebral

Hui-Yang Huang

Yohannes Tsehay

Francesca Mertan

Dave A Roy

Yifan Peng

Daniel Joseph Mollura

Zhiyong Lu

A new chest X-ray database, referred to as 'ChestX-ray8', is disclosed herein, which comprises over 100,000 frontal view X-ray images of over 32,000 unique patients with the text-mined eight disease image labels (where each image can have multi-labels), from the associated radiological reports using natural language processing. We demonstrate that these commonly occurring thoracic diseases can be detected and spatially-located via a unified weakly supervised multi-label image classification and disease localization framework, which is validated using our disclosed dataset.

Method and system of building hospital-scale chest X-ray database for entity extraction and weakly-supervised classification and localization of common thorax diseases

Disclosed prostate computer aided diagnosis (CAD) systems employ a Random Forest classifier to detect prostate cancer. System classify individual pixels inside the prostate as potential sites of cancer using a combination of spatial, intensity and texture features extracted from three sequences. The Random Forest training considers instance-level weighting for equal treatment of small and large cancerous lesions and small and large prostate backgrounds. Two other approaches are based on an AutoContext pipeline intended to make better use of sequence-specific patterns. Also disclosed are methods and systems for accurate automatic segmentation of the prostate in MRI. Methods can include both patch-based and holistic (image-to-image) deep learning methods for segmentation of the prostate. A patch-based convolutional network aims to refine the prostate contour given an initialization. A method for end- to-end prostate segmentation integrates holistically nested edge detection with fully convolutional networks. HNNs automatically learn a hierarchical representation that improve prostate boundary detection.

Detection of prostate cancer in multi-parametric MRI using random forest with instance weighting and MR prostate segmentation by deep learning with holistically-nested networks

Methods include processing image data through a plurality of network stages of a progressively holistically nested convolutional neural network, wherein the processing the image data includes producing a side output from a network stage m, of the network stages, where m&#x3e;1, based on a progressive combination of an activation output from the network stage m and an activation output from a preceding stage m&#x2212;1. Image segmentations are produced. Systems include a 3D imaging system operable to obtain 3D imaging data for a patient including a target anatomical body, and a computing system comprising a processor, memory, and software, the computing system operable to process the 3D imaging data through a plurality of progressively holistically nested convolutional neural network stages of a convolutional neural network.

Progressive and multi-path holistically nested networks for segmentation

Methods and systems for diagnosing cancer in the prostate and other organs are disclosed. Exemplary methods comprises extracting texture information from MRI imaging data for a target organ, sometimes using two or more different imaging modalities. Texture features are determined that are indicative of cancer by identifying frequent texture patterns. A classification model is generated based on the determined texture features that are indicative of cancer, and diagnostic cancer prediction information for the target organ is then generated to help diagnose cancer in the organ.

Automated cancer detection using MRI

An image of an anatomical structure can be analyzed to determine an enclosing three-dimensional boundary when the anatomical structure is filled with two substances, such as air and a fluid. Various techniques can be used to determine the enclosing boundary including: analyzing the virtual structure to segment the structure into air and fluid pockets, determining if there are multiple fluid pockets whose surface touches a single air-fluid boundary, determining a separate threshold for respective fluid pockets, resegmenting the virtual anatomical structure using the separate threshold for different fluid pockets, forming a hierarchical pocket tree which represents the relationship between the fluid and air pockets, pruning the pocket tree based on various criteria which corresponds to deleting those pruned portions from the virtual anatomical structure, and resegmenting the remaining virtual anatomical structure using one or more of fuzzy connectedness, two-dimensional gap filling, and level set segmentation.

Hybrid segmentation of anatomical structure

Candidate anomalies in an anatomical structure are processed for classification. For example, false positives can be reduced by techniques related to the anomaly's neck, wall thickness associated with the anomaly, template matching performed for the anomaly, or some combination thereof. The various techniques can be combined for use in a classifier, which can determine whether the anomaly is of interest. For example, a computed tomography scan of a colon can be analyzed to determine whether a candidate anomaly is a polyp. The technologies can be applied to a variety of other scenarios involving other anatomical structures.

Computer-aided classification of anomalies in anatomical structures

Various level set techniques can be used to automatically segment the colon wall, including identifying the colon wall outer boundary. A speed image can be used during level set processing. For example, the speed image can be generated via inverting the gradient perpendicular to the segmented inner boundary of the colon wall. The techniques can be useful for determining wall thickness, which can be used to classify polyp candidates, diagnose diseases of the colon, and the like.

Segmenting colon wall via level set techniques

Various techniques can be used to improve classification of colon polyps candidates found via computed tomographic colonography computer aided detection (CTCCAD). A polyp candidate can be classified as a true positive or a false positive. For example, a two-dimensional projection image of the polyp can be generated from a three-dimensional representation and classified based on features of the projection image. An optimal viewpoint for the projection image can be found via techniques such as maximizing viewpoint entropy. Wavelet processing can be used to extract features from the two-dimensional projection image. Feature extraction can use a piecewise linear orthonormal floating search for locating most predictive neighbors for wavelet coefficients, and support vector machines can be employed for classification. The techniques can be useful for improving accuracy of CTCCAD techniques.

Virtual colonoscopy via wavelets

A three dimensional image of the colon like surface is processed to determine at least its ring structure. The image is composed of vertex points, each vertex point having a discrete point identifier and three dimensional position information. The three dimensional position information is averaged in a shrinking procedure to contract the three dimensional image proximate to a major axis of the colon-like surface. Evenly spaced points are taken through the shrunken colon like surface and connected to form a curve. Planes are generated at intervals normal to the curve to split the shrunken colon like surface into image segments. By mapping these image segments back to the original image through their discrete point identifiers, an accurate ring profile of the colon like surface can be generated.

Automated centerline detection algorithm for colon-like 3D surfaces

A computer-assisted method for detecting surface features in a virtual colonoscopy. The method includes providing a three-dimensional construction of a computed tomography colonography surface; creating a path along the teniae coli from the proximal ascending colon to the distal descending colon on the colonography surface; forming an indexed computed tomography colonography surface using the created path; and registering the supine and prone scans of the computed tomography colonography surface using the indexed computed tomography colonography surface. The method also includes navigating the internal surface of the computed tomography colonography using the indexed computed tomography colonography surface.

Teniae coli guided navigation and registration for virtual colonoscopy

A virtual anatomical structure can be analyzed to determine enclosing three-dimensional boundaries of features therein. Various techniques can be used to determine tissue types in the virtual anatomical structure. For example, tissue types can be determined via an iso-boundary between lumen and air in the virtual anatomical structure and a fuzzy clustering approach. Based on the tissue type determination, a deformable model approach can be used to determine an enclosing three-dimensional boundary of a feature in the virtual anatomical structure. The enclosing three-dimensional boundary can be used to determine characteristics of the feature and classify it as of interest or not of interest.

Determination of feature boundaries in a digital representation of an anatomical structure

Portions of a virtual colon can be analyzed to identify a normal structure, such as an ileocecal valve. Paradigmatic characteristics of ileocecal valves can be used to identify a digital representation as an ileocecal valve. Upon determination that a digital representation has the characteristics of an ileocecal valve, action can be taken. For example, the digital representation can be removed from a list of polyp candidates.

Automated identification of ileocecal valve

Growing community of inventors

Potomac, MD, United States of America

Ronald M Summers