AUTOMATIC SEGMENTATION OF LARYNGEAL CARTILAGES USING SUPPORT VECTOR MACHINES

R. Reeve Ingle1, Berhane H. Azage1, Joëlle K. Barral1, Kie Tae Kwon1, Edward G. Damrose2, Nancy J. Fischbein2,3, Dwight G. Nishimura1

1Electrical Engineering, Stanford University, Stanford, CA, 2Otolaryngology, Stanford University, Stanford, CA, 3Radiology, Stanford University, Stanford, CA

Introduction Results

Conclusions

Methods

The presence and extent of cartilage invasion by laryngeal carcinoma is difficult to assess on routine MR imaging studies. The ultimate goal of this work is to apply image segmentation to assess the extent of laryngeal cartilage invasion by tumor. While fully-automated segmentation of the laryngeal cartilages remains unexplored, a multi-contrast and multi-dimensional approach has proven useful for segmenting articular cartilage [1]. This approach is hindered by a lack of automatic intensity cor-rection to compensate for the coil sensitivity profile when a dedicated array is used.

We propose a custom intensity correction algorithm, and we explore the use of supervised and unsupervised learning algorithms to automatically segment the car-tilages from high-resolution larynx images of healthy volunteers. We compare the performance of two methods: (1) support vector machines (SVMs), a supervised learning algorithm that uses training data to build a classification model, and (2) k-means clustering, an unsupervised learning algorithm (requiring no training) that classifies input data by grouping datapoints with similar features into k clusters.

References[1] Koo et al. In: ISMRM, Toronto, Canada, 2008, p. 2546.[2] Barral et al. In: ISMRM, Honolulu, HI, 2009, p. 1318.[3] http://white.stanford.edu/software/.

[4] Gering et al. JMRI, 13(6):967-975, 2001.[5] Chang et al. LIBSVM, http://www.csie.ntu.edu.tw/~cjlin/libsvm.[6] Styner et al. IEEE Trans Med Imaging, 19(3):153-165, 2000.

Scans were conducted on a 1.5 T GE Signa scanner using a larynx-dedicated three-channel array [2]. Four 2D multi-slice sequences were used on healthy vol-unteers, with the following scan parameters: FOV = 10 cm, BW = ±32 kHz, resolu-tion = 0.78 x 0.39 mm2, slice thickness = 2 mm, # averages = 2. Timing parameters specific to each sequence are listed in Table 1, and the resulting images from one slice of a larynx dataset are shown in Fig. 1.

We have successfully applied the SVM and k-means algorithms to segment the cartilages from MR images of the larynx. The implementation of an intensity correc-tion technique significantly improved the performance of automatic segmentation. k-means clustering produced comparable segmentation results to SVM, with a slight improvement in overall accuracy and accuracy of cartilage classification. The main advantage of k-means over SVM classification is that it requires no manual segmen-tation of training data, which is tedious, time-consuming, and subject to error.

Preprocessing: • Registration performed automatically

using mrVista [3]• Intensity correction to compensate for

uneven coil sensitivity profile Figure 2 describes the intensity correction algorithm and shows one slice of the PD dataset before and after intensity correction.Image segmentation via SVM:• Training data manually segmented

using 3DSlicer [4]• LibSVM used for multi-class SVM [5]

Image segmentation via k-means:• Carried out for k=3 clusters• Repeated four times to avoid

local minima

MR RL

Figure 1. Larynx images

Images from one slice of a larynx data-set. Four MR sequences were chosen to yield different contrasts:

(a) Proton-density-weightedspin echo (PD)

(b) Spin echo (SE)(c) Fast spin echo with IDEAL

water-fat separation (FSE-IDEAL)(d) Fast spin echo XL (FSE-XL)

a) b)

c) d)

20008003500800

FSE-IDEALFSE-XL

PDSE

TR (ms)Sequence20321018

TE (ms)Table 1. Imaging parameters

Con

trast

3

Contrast 2Contrast 1

Each pixel in the manually-segmented test-ing dataset is plotted according to the in-tensities of contrast 1 (PD), contrast 2 (SE), and contrast 3 (FSE-IDEAL). Pixels are classified as:• Cartilage and fat (yellow)• Muscle (red)• Trachea and background (black)Well-localized clusters suggest that automatic segmentation will achieve good accuracy.

Figure 3. Scatter plot of pixel intensities

PD Image

ManuallySegmented

“Gold Standard”

SVMLabeling

k-meansClustering

Slice 1 Slice 2 Slice 3 Slice 4 Slice 5

SVM and k-means testing results are shown for five test slices from a larynx dataset. SVM training and testing data came from two different subjects. Re-sults of SVM labeling demonstrate classification of cartilage and fat (yellow), muscle (red), and trachea and background (white) with 93% accuracy w.r.t. manual segmentation. Among the cartilage pixels, 69% are classified correctly by SVM. k-means clustering results in comparable segmentation and achieves an overall accuracy of 94% (77% for cartilage pixels) without requiring te-dious manual segmentation.

Figure 5. Image segmentation results

CONTACT: RINGLE@STANFORD.EDU

a) b)

c) d)

(a) Original PD image(b) Intensity-corrected image(c) Mask used in polynomial fitting(d) Resulting third-order polynomial fit

Algorithm: low-order polynomial fit [6]minimize ||W(Xa - y)||2subject to Xa > 0

W = diag. matrix of binary mask weights X = matrix of image coord. monomials

(xnym) up to desired fitting order y = vector of original image values a = vector of polynomial coefficients

Figure 2. Intensity correction

Results of SVM (top row) and k-means (bottom row) on an uncorrected image (left column) and the same image after intensity correction (right column). Due to the uneven coil sensi-tivity profile, image intensity varies with posi-tion, and intensity correction significantly im-proves segmentation results.

Figure 4. Effects of intensity correction on image segmentation

Uncorrected Corrected

SVM

k-m

eans