Article
A computer-based approach for classification of cervical spine defects in MR images
Search Medline for
Authors
-
Stefan Daenzer
- Innovationszentrum für Computerassistierte Chirurgie, Universität Leipzig
-
Stefan Freitag
- Innovationszentrum für Computerassistierte Chirurgie, Universität Leipzig
-
Sandra v. Sachsen
- Innovationszentrum für Computerassistierte Chirurgie, Universität Leipzig
-
Hanno Steinke
- Institut für Anatomie, Universitätsklinikum Leipzig
-
Mathias Groll
- Klinik für Neurochirurgie, Universitätsklinikum Leipzig
-
Jürgen Meixensberger
- Innovationszentrum für Computerassistierte Chirurgie, Universität Leipzig
-
Mario Leimert
- Klinik für Neurochirurgie, Universitätsklinikum Carl Gustav Carus Dresden
| Published: | May 13, 2014 |
|---|
Outline
Text
Objective: Fully automatic computer-based classification of spinal defects in Magnetic Resonance (MR) images, such as cervical spine stenosis, is an important step towards automatic diagnosis assistance systems. However, defect classification is a challenging task due to natural and pathological anatomical variations, varying MR acquisition parameters and inhomogeneities in image contrast. In this work we introduce a fully automatic algorithm for classification of cervical spine stenosis in MR images.
Method: We present a novel machine learning method for cervical spine stenosis classification based on features derived from a MR image in combination with a linear support vector machine (SVM). A novel method for bivariate gradient orientation histogram generation from 3D raster image data is used. This allows us to describe a stenosis using bivariate histograms. We generate a window containing the expected stenosis area from the center of gravity (CoG) of two adjacent vertebrae using simple trigonometric calculations and translations. The first step is to calculate the center between two vertebrae CoGs. From this center point an orthogonal translation is performed in direction of the spinal canal. A box with predefined size is then generated at this position. This area of the image is resampled to a predefined resolution to serve as input for the stenosis gradient orientation feature generation.
Results: The performance of the developed algorithm was evaluated on 20 T2-weighted MR images of the cervical spine. We have manually annotated four regions (between C3–C7) per MR image individually as “Stenosis” and “No Stenosis”. In a leave-one-out study on our MR image dataset the proposed algorithm reached a classification accuracy of 82.5 percent on 80 samples (57 positive, 23 negative samples). The classifier achieved a sensitivity of 0.93, and a specificity of 0.57 in our study.
Conclusions: We have introduced a fully automatic computer-based classification method for cervical spine stenosis. The proposed method achieved results for stenosis classification in MR images which makes the usage of the algorithm feasible in a broad spectrum of diagnosis related clinical applications. The new method represents an important step towards computer-based automatic diagnosis of spinal stenosis without the need of clinical expert interaction.
