Machine Learning-based Human Knee Cartilage Segmentation on MRI
Siddhi Munde1, Melissa N Manzer1, Wellsandt Elizabeth2, Jessica Emory3, and Balasrinivasa R Sajja1

1Radiology, University of Nebraska Medical Center, Omaha, NE, United States, 2Division of Physical Therapy, University of Nebraska Medical Center, Omaha, NE, United States, 3University of Nebraska Medical Center, Omaha, NE, United States


Accurate knee cartilage segmentation on MRI is essential to obtain quantitative measures from cartilage that help in the assessment of knee pathology and therapeutic response in patients with diseases such as Osteoarthritis. Segmentation of cartilage on routine clinical MRI is challenging due to image intensity variation across the structure and low image contrast. In this study, we obtained an accurate cartilage segmentation on PD and T1 weighted images using Support Vector Machine (SVM) classifier with a spatial indexing feature which accounts for regional signal variations.


Tracking morphological changes in knee cartilage provide valuable information to clinicians to assess disease progression and response to treatment in patients suffering from diseases such as Osteoarthritis. Consequently, it is essential to have a robust method to accurately segment knee cartilage on MRI. The variability in structure and intensities of cartilage along with low image contrast hinders accurate automated segmentation methods. This becomes even more challenging if only routine clinical MRI are used for segmentation. To this end, we have implemented a Support Vector Machine based classification method with a new feature of grid based spatial indexing and have successfully segmented cartilage on routine clinical MRI.

Materials and Methods

MRI Data: Proton density (PD) and T1 weighted (T1w) MRI were acquired on a 3T scanner using a 15-channel knee coil with image parameters: image size=320x320, orientation=Sagittal, acquisition type=2D, FOV=140mmx140mm, number of slices=28, slice thickness=3 mm, spacing between slices=3.75 mm. For PD with fat saturation pulse, TR/TE=2430/26 ms. For T1w, TR/TE=642/18 ms. Subjects: Nine volunteers with no history of knee injury or pain were included in this study and informed consent was obtained from them. Manual segmentations of cartilage for all datasets were performed by a trained person with segmentations confirmed by a board-certified musculoskeletal radiologist. Data from four randomly selected subjects were used for training the Support Vector Machine (SVM) classifier. Predictions from SVM were obtained on the remaining 5 subjects. Figure 1 shows the schematic representation of various steps involved in preprocessing, feature extraction, SVM training and classification. Pre-processing: Preprocessing was performed on all nine participants’ data. (1) PD and T1w rigid-body registered to correct for between sequences movement. (2) Image bias field corrected with N3 method. (3) Intensity was standardized so cartilage will have similar intensities from all subjects. (4) Image thresholded with Otsu’s method to separate foreground from background. Feature vector consisted of cartilage intensities from PD, T1w images, distance from bone boundary, and grid based spatial index. Distance from bone: To measure distance from bone, edges of femur and tibia bone were extracted from anisotropic diffusion filtered PD image by Canny edge detector. Basing on connectivity and segments’ size, small edges were removed. Euclidean distance map was computed from bone boundary. Automated grid placement: A new feature of grid based spatial index over cartilage region in the image was introduced. It distinctly adds the spatial reference of the cartilage with respect to femur bone. It was to done determine the location and corresponding intensities of different parts of the cartilage for SVM classifier. The maximum width and lowest surface of the femur bone was identified from center 5 slices of the bone boundary to generate a box. The box size was automatically determined so it completely covered cartilage on each slice. Horizontal numbering as index for identifying every grid was defined as a descriptor for classification. SVM Classifier: A supervised machine learning method, SVM with radial basis function (RBF) kernel from R package library e1071 was used as a discriminative classifier. For training, feature vectors with class labels from manual segmentation were used. The classification results on five subjects were quantitatively evaluated with Dice Similarity Coefficient (DSC), sensitivity, and specificity by comparing with manually segmented cartilages.


In Figure 2, 2(A) and 2(B) show representative PD and T1w images. 2(C) and 2(D) show the detected bone edge and Euclidean distance map. 2(E) shows the adaptive 16 grid placement for spatial indexing. The SVM segmented cartilage is shown in 2(F). Quality of the segmentation can be appreciated on this image. Table 1 shows quantitative assessment of SVM cartilage prediction results on all five test subjects. The quality of segmentation has increased with increase in number of grids. The chart from Figure 3 shows a linear increase in %DSC with number of grids compared to predictions without grid. A percentage increase of 11.8% up to 16 grids and no significant improvement beyond is observed. The segmentation results with the present method using 16 grids demonstrated the optimal segmentation results.


We have presented a Support Vector Machine based method that applied on routine clinical PD and T1w MRI for segmenting human knee cartilage. We have introduced a new feature of spatial indexing over complete cartilage region into SVM training and classification. A 4x4 grid is found to be optimal in this study. This feature along with distance to the bone has significantly improved the quality of segmentation, within feasible computational time. The segmentation can be further improved if data from specific MRI sequences such as DESS are used.


No acknowledgement found.


1. Zhang K, Lu W, and Marziliano. (2013). Automated knee cartilage segmentation from multi-contrast MR images using support vector machine classification with spatial dependencies. Magn. Reson. Imag. 31, 1731-1743.

2. Paul, P. K., Kris Jasani, M., Sebok, D., Rakhit, A., Dunton, A. W., & Douglas, F. L. (1993). Variation in MR signal intensity across normal human knee cartilage. Journal of magnetic resonance imaging, 3(4), 569-574.

3. R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/.


Figure 1: Schematic representation of various steps involved in pre-processing and feature extraction for SVM training and predictions.

Figure 2: A and B: Representative images of PD and T1w. C: Overlay of detected edge as bone boundary. D: Distance map computed from bone boundary. E: Overlay of automated 16 grid placement and spatial indexing. Values from A, B, D, E form feature vectors for SVM training and classification. F: Overlay of segmented cartilage obtained from SVM predictions.

Table 1: Mean ± standard deviation values of Dice Similarity Coefficient (DSC), sensitivity and specificity of predicted cartilage segmentation using Support Vector Machine (SVM) method on all five subjects. Accuracy of segmentation increased with the number of grids.

Figure 3: Percentage increase in Dice Similarity Coefficient (DSC) with the increase in number of grids compared to DSC with no grids.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)