US-12620088-B2 - System and method for articular cartilage thickness mapping and lesion quantification

US12620088B2US 12620088 B2US12620088 B2US 12620088B2US-12620088-B2

Abstract

Systems and methods for articular cartilage thickness mapping and lesion quantification operate on 3D medical image data to reconstruct cartilage surfaces, estimate surface normals, determine cartilage thickness, and identify regions of full-thickness cartilage loss (FCL). Reconstructed cartilage surfaces can be parcellated into subregions using a rule-based approach.

Inventors

Yongcheng YAO
Weitian Chen

Assignees

THE CHINESE UNIVERSITY OF HONG KONG

Dates

Publication Date: 20260505
Application Date: 20230110

Claims (13)

1 . A method of analyzing image data including image data corresponding to a region of cartilage, the method comprising: constructing, using the image data, a first surface mesh corresponding to a first cartilage surface and a second surface mesh corresponding to a second cartilage surface, wherein each of the first surface mesh and the second surface mesh comprises a plurality of vertices and wherein constructing the first surface mesh and the second surface mesh includes: identifying, in the image data, a set of voxels corresponding to the region of cartilage; extracting, from the set of voxels, a first subset of voxels corresponding to an inner surface of cartilage and a second subset of voxels corresponding to an outer surface of cartilage; defining the first surface mesh from the first subset of voxels and the second surface mesh from the second subset of voxels; and performing a surface closing operation on the first surface mesh, wherein the surface closing operation includes one or more iterations of a surface dilation operation on the first surface mesh and one or more iterations of a surface erosion operation on the first surface mesh, wherein each iteration of the surface dilation operation adds neighboring vertices of the region of cartilage to the first surface mesh and each iteration of the surface erosion operation removes marginal vertices from the first surface mesh; estimating surface normals corresponding to a plurality of locations on the first surface mesh; computing a distance along each surface normal from the first surface mesh to the second surface mesh; generating a three-dimensional (3D) cartilage thickness map for the region of cartilage based at least in part on the computed distances; and rendering the 3D cartilage thickness map on a display device.
2 . The method of claim 1 wherein estimating the surface normal includes, for each location in the plurality of locations on the first surface mesh: identifying a plurality of neighbor locations to the location; constructing a centered position matrix for the location and the plurality of neighbor locations; performing a singular value decomposition on the centered position matrix; and extracting an initial surface normal from the singular value decomposition.
3 . The method of claim 2 wherein estimating the surface normals further includes: after extracting the initial surface normal for each location in the plurality of locations, performing orientation correction on the initial surface normals such that the surface normals are oriented toward the second surface mesh.
4 . A method of analyzing image data including image data corresponding to a region of cartilage, the method comprising: constructing, using the image data, a first surface mesh corresponding to a bone- cartilage interface surface, wherein the first surface mesh comprises a plurality of vertices and wherein constructing the first surface mesh includes: identifying, in the image data, a set of voxels corresponding to the region of cartilage; extracting, from the set of voxels, a first subset of voxels corresponding to an inner surface of the region of cartilage; defining the first surface mesh from the first subset of voxels; and performing a surface closing operation on at least the first surface mesh, wherein the surface closing operation includes one or more iterations of a surface dilation operation on the first surface mesh and one or more iterations of a surface erosion operation on the first surface mesh, wherein each iteration of the surface dilation operation adds neighboring vertices of the region of cartilage to the first surface mesh and each iteration of the surface erosion operation removes marginal vertices from the first surface mesh; constructing, using the image data and a cartilage segmentation template, a second surface mesh corresponding to a complete bone-cartilage interface surface; computing, based on the first surface mesh and the second surface mesh, a third surface mesh representing a region of full-thickness cartilage loss (FCL); and rendering, on a display device, an FCL map based on the second and third surface meshes.
5 . The method of claim 4 further comprising: applying a plurality of parcellation rules to define a plurality of sub-regions of the first surface mesh; and computing a percentage of FCL for each sub-region based on the third surface mesh and the second surface mesh.
6 . The method of claim 4 wherein constructing the second surface mesh includes: obtaining, using the image data, a bone mask and a cartilage mask; defining a bone surface mesh using the bone mask; determining a deformation field from the image data; applying the deformation field to a template cartilage segmentation mask representing normal cartilage to generate a warped cartilage segmentation; combining the warped cartilage segmentation with the cartilage mask to produce the second surface mesh; and refining the second surface mesh.
7 . The method of claim 6 wherein refining the second surface mesh includes applying one or more surface-hole-filling operations to the second surface mesh.
8 . The method of claim 7 wherein the one or more surface-hole- filling operations include a connectivity-based hole-filling operation followed by a curve-based hole filling operation.
9 . A method of analyzing image data including image data corresponding to a region of cartilage, the method comprising: extracting, from the image data, a first set of voxels for an inner cartilage surface corresponding to a bone-cartilage interface; extracting, from the image data, a second set of voxels for an outer cartilage surface; defining an inner surface mesh from the first set of voxels and an outer surface mesh from the second set of voxels, wherein each of the inner surface mesh and the outer surface mesh comprises a plurality of vertices; and refining the inner surface mesh and the outer surface mesh, wherein refining the inner surface mesh and the outer surface mesh includes performing one or more iterations of a restricted surface dilation operation on the outer surface mesh, wherein each iteration of the restricted surface dilation operation adds neighboring vertices from the region of cartilage, exclusive of the inner surface mesh, to the outer surface mesh; and rendering a three-dimensional (3D) image of the region of cartilage on a display device, based at least in part on the inner surface mesh and the outer surface mesh.
10 . The method of claim 9 wherein refining the inner surface mesh and the outer surface mesh includes: performing a surface closing operation on the inner surface mesh.
11 . The method of claim 10 wherein performing the surface closing operation includes: performing one or more iterations of a surface dilation operation on the inner surface mesh, wherein each iteration of the surface dilation operation adds neighboring vertices of the region of cartilage to the inner surface mesh; thereafter performing one or more iterations of a surface erosion operation on the inner surface mesh, wherein each iteration of the surface erosion operation removes marginal vertices from the inner surface mesh.
12 . The method of claim 9 further comprising: obtaining a cartilage mask identifying one or more voxels in the image data as corresponding to cartilage; constructing a skeleton surface using the image data; estimating a surface normal for each voxel on the skeleton surface; and selecting, for each surface normal, a voxel identified in the cartilage mask as one of the first set of voxels.
13 . The method of claim 1 wherein computing the distance along each surface normal from the first surface mesh to the second surface mesh further comprises, for at least one of the surface normals: computing an initial distance along the surface normal from the first surface mesh to the second surface mesh; determining an uncertainty associated with the surface normal; and adding the uncertainty to the initial distance.

Description

CROSS-REFERENCES TO RELATED APPLICATIONS This application claims the benefit of U.S. Provisional Application No. 63/266,663, filed Jan. 11, 2022, the disclosure of which is incorporated by reference herein. TECHNICAL FIELD The present disclosure relates generally to analysis of medical images of a joint (such as a knee) and in particular to systems and methods for articular cartilage thickness mapping and lesion quantification. BACKGROUND Osteoarthritis (OA) is one of the main causes of disability in older individuals. Efforts have been made to study the pathology of OA from various aspects. One area of interest is the progressive and irreversible cartilage loss at different clinical stages of OA. Early studies investigated the relationship between cartilage changes and disease severity using the mean thickness values of cartilage in certain pre-defined regions of interest (ROIs). Using mean thickness values of ROIs to correlate thickness changes to OA, while simple, omits important information on spatial variation of the thickness. Research has shown, for instance, that the spatial variation pattern of femoral cartilage thickness can facilitate the detection of OA-related differences. In recent years, research efforts that focus on the cartilage thickness pattern and utilize all information from the thickness map have been reported. Regardless of which scheme is adopted, the first and foremost step is accurate cartilage thickness measurement. Existing techniques for cartilage thickness measurement include the nearest neighbor approach, the surface normal approach, the local thickness approach, and the potential field line approach. Among these, the nearest neighbor and surface normal approaches are the most widely applied methods. A nearest neighbor approach can be implemented using a brute force search algorithm that can calculate the Euclidean distance for each pair of points from two point-sets and find the shortest one. Alternatively, a sphere growing algorithm, which is generally more efficient, can be used. For the surface normal approach, one challenge is estimating the normal vector. Various studies have estimated a two-dimensional (2D) surface normal by fitting a section of curve to a cubic equation and estimated three-dimensional (3D) surface normal using principal component analysis (PCA). In general, cartilage thickness measurement can be carried out in 2D or 3D space. Three-dimensional measurement has advantages, because 2D assessment heavily depends on the scanning position and is more likely to produce erroneous thickness values. However, 3D measurement approaches are less well developed. SUMMARY Certain embodiments of the present invention relate to systems and methods for articular cartilage thickness mapping and lesion quantification. Such systems and methods have applications in diagnosis and monitoring of osteoarthritis and other medical conditions. Systems and methods described herein can be applied to medical images obtained using various imaging modalities (including but not limited to magnetic resonance imaging) that can be used to generate image data including voxels, where each voxel corresponds to a discrete region in 3D space. Using the image data, a bone segmentation mask and a cartilage segmentation mask can be defined that associates particular voxels with tissue of a particular type (e.g., bone or cartilage). It should be understood that a particular image segmentation technique is not required and that the techniques described herein for mapping and measurement of cartilage thickness can be applied to any segmented 3D images for which a cartilage mask and/or a bone mask have been generated. According to some embodiments, the segmentation masks can be defined using machine learning techniques. For instance, a first deep learning network (which can be, e.g., a convolutional neural network or the like) can be trained (e.g., based on training data that includes images labeled to indicate ground truth cartilage regions) to perform image segmentation on three-dimensional image data to generate a cartilage mask and optionally a bone mask. Jointly with training the first deep learning network, a second deep learning network (which can also be, e.g., a convolutional neural network or the like) can be trained to perform registration between a segmented image and a template image indicating a normal arrangement of cartilage in the target joint, e.g., by determining a deformation field to apply to the template that reshapes the template to match the segmented image. In some embodiments, the template image can be learned jointly with training of the first and second deep learning networks. For example, the template image can be initialized based on an average of the training data and updated during training. After training, the first and second deep learning networks can be applied to testing data, which can include 3D image data generated by imaging the target joint in a test subject.