Standalone annotations of axial-view spine images
AssigneesLondon Health Sciences Centre Research Inc., General Electric Company
InventorsIsmail Ben Ayed, Brandon Miles, Gregory Garvin
Life of this patent
Estimating the term…
Abstract
Systems, methods and computer program products to annotate axial-view spine images with the desired characteristics of not requiring additional views of the spine or cross-referencing features are provided. In one aspect, the disclosed method does not require external training from a manually-labeled data set while being applicable to different imaging modalities and acquisition protocols. In one aspect, the disclosed method achieves near real-time results using integral kernels when implemented on a Graphics Processing Unit.
Claims10
A computer-implemented method to create annotations of spine images, the method comprising:
- loading reference data comprising prior geometric measurements of the human spine;
- receiving an initial input from the selection of a point within a vertebral region in a single, two-dimensional axial-view image slice of a spinal magnetic resonance image series;
- generating a target distribution of identified pixel intensity values for said vertebral region;
- classifying each of a plurality of pixel intensity values of each said image slice of said spinal image series to determine whether such pixel intensity value belongs to vertebrae by matching each pixel context to said target distribution using a non-linear probability product integral kernel;
- wherein the context of the pixel is the kernel density estimate of intensities within a window centered at the pixel;
- classifying each of said image slices of said spinal image series by computing a two-dimensional slice-level feature which is derived from the classification of said plurality of pixels of said spinal image series and from said reference data;
- identifying three-dimensional vertebra from said two-dimensional image slice-level features;
- and
- generating an annotation to at least one of said two-dimensional axial image slices identifying at least one vertebra or an associated intervertebral disk.
A computer storage device including program instructions for execution by a computing device to perform:
- loading reference data comprising prior geometric measurements of the human spine;
- receiving an initial input from the selection of a point within a vertebral region in a single, two-dimensional axial-view image slice of a spinal magnetic resonance image series;
- generating a target distribution of identified pixel intensity values for said vertebral region;
- classifying each of a plurality of pixel intensity values of each said slice of said spinal image series to determine whether such pixel intensity value belongs to vertebrae by matching each pixel context to said target distribution using a non-linear probability product integral kernel;
- wherein the context of the pixel is the kernel density estimate of intensities within a window centered at the pixel;
- classifying each of said image slices of said spinal image series by computing a two-dimensional image slice-level feature which is derived from the classification of said plurality of pixels of said spinal image series and from said reference data;
- identifying three-dimensional vertebra from said two-dimensional image slice-level features;
- and
- generating an annotation to at least one of said two-dimensional axial image slices identifying at least one vertebra or an associated intervertebral disk.
Dependencies are inferred from the claim text.