Texture-Based Detection of Myositis in Ultrasonographies Tim König 1, Marko Rak 1, Johannes Steffen 1, Grit Neumann 2, Ludwig von Rohden 2, Klaus D. Tönnies 1 1 Institut für Simulation & Graphik, Otto-von-Guericke-Universität Magdeburg 2 Klinik für Radiologie & Nuklearmedizin, Otto-von-Guericke-Universität Magdeburg tim.koenig@st.ovgu.de Abstract. Muscle ultrasonography is a convenient technique to visualize healthy and pathological muscle tissue as it is non-invasive and image acquisition can be done in real-time. In this paper, a texture-based approach is presented to detect myositis in ultrasound images automatically. We compute different texture features like wavelet transform features and first-order grey-level intensity statistics of a relevant central image patch carrying structure and intensity information of muscle tissue. Using a combination of these information we reached an accuracy of classification of 92.20 % with our approach on a training data set of 63 clinically pre-classified data sets. 1 Introduction An automatically computer-supported detection of myositis may serve different purposes and goals: it can save time for radiologists during the diagnosis and it gives a second, independent result, which can be taken into consideration. A well trained process may also be capable of early diagnosis and categorization. Moreover, it will help young professionals with only a small degree of experience identifying myositis in ultrasound images. Neuromuscular disorders often cause structural muscle changes that can be seen in ultrasound images. Infiltration of fat and fibrous tissue increase muscle echo intensity, i.e. the reference image of the muscle will become brighter. Thus, myositis can be found by measuring muscle thickness. As healthy muscles contain only little fibrous tissue only a few reflections will occur during image acquisition resulting in a low echo intensity and thus a relatively dark image (Fig. 1). It is assumed that the replacement of muscle tissue with fat and fibrosis is the main cause of increased muscle echo intensity as they increase the number of reflections within the muscle and therefore the mean grey value of the muscle in the ultrasound image (Fig. 1)[1]. Under these assumptions a texture-based analysis should be adequate to detect myositis. Using texture as discriminating feature in ultrasound images has a long history in image analysis. Most methods use a combination of spectral and first-order statistical features as these reflect the nature of deterministic H.-P. Meinzer et al. (Hrsg.), Bildverarbeitung für die Medizin 2013, Informatik aktuell, DOI: 10.1007/978-3-642-36480-8_16, Springer-Verlag Berlin Heidelberg 2013 81
82 König et al. tissue-related variation and non-deterministic influences from image generation (e.g. [2] for classification of liver tumors, [3] for classification of breast tumors, or [4] for the classification of arterioscleroric tissue). Both, differences in grey-level intensity and in the micro-structure between healthy and pathological muscle tissue can be used to compute representative and distinctive features. In contrast to [5], who required training to select texture features and a manually specified region-of-interest (ROI), we believe that - similar to [4] - selection of features reflecting micro-structural change will result in a robust method that requires little to no user input. Fig. 1 shows that the texture of pathological muscle tissue images is more unstructured or diffuse, containing smaller contrast changes compared to the texture of healthy muscle tissue images, which seem to be more directed or structured containing higher contrast changes caused by the relatively dark regions of the healthy muscles. We propose a texture analysis method that applies a wavelet decomposition computing features carrying structure information as well as first-order grey-level intensity statistics. Fig. 1. Ultrasound images including healthy (left) and pathological (right) muscle tissue. 2 Materials and methods Feature extraction and classification are performed on 63 clinical ultrasound images including 14 images of healthy muscles and 49 images of pathological muscle tissues. The images have a pixel resolution of 350 px 500 px covering an area of 2.8 cm 4.0 cm. The data contains scans of various muscles from upper and lower human limbs as well as scans acquired by different angulation of the 9MHz linear transducer. All images were pre-classified by radiological experts allowing a comparison of our analysis results with these ground truth assumptions.
Detection of Myositis in Ultrasonographies 83 2.1 Feature extraction In the feature extraction, we focus on two features of the wavelet decomposition and two features of the first-order grey-level intensity statistics: the reverse biorthogonal 1.1 and 1.3 wavelet transform features [6] carrying non-directional information of the muscle tissue micro-structure and also the grey-level entropy and variance as first-order statistic features. These features measure similar attributes compared to [2] and especially [4]. They were selected from a range of features by visual inspection of scatter plots of all possible pairs of features. Since the data does not include a segmentation of the muscle tissue and a manual search of a window for which the features are computed would be inefficient we specified during image acquisition that the relevant muscle structures are located near the center of the ultrasound image. Hence, we place a ROI of size 60 px 60 px in the center of the image and compute the four texture features for each pixel p in the ROI (Fig. 2). Texture computation requires the definition of a window with suitable size around p. We experimented with window sizes between 80 px 80 px and 220 px 220 px for determining the optimal size. Fig. 2. Feature extraction for each pixel of a central ROI (red) of 60 px 60 px with different window sizes (yellow) from 80 px 80 px up to 220 px 220 px. This results in 3600 feature vectors per image for a fixed window size. Each feature vector contains the four different features. 2.2 Classification To classify the images we use the k-nearest neighbor algorithm (knn) [7] based on closest training samples in feature space. The knn classification is then crossvalidated by the leave-one-out method. Thus, the knn classifier was trained on the feature subset of N 1 images each image contains 3600 feature vectors for a fixed windows size to validate the class of the N-th image (with N = 63). This process is repeated such that each image is used once for validation. However, the choice of k is highly dependent on the problem and the data set. Hence, we experimented with k varying from 2 up to 128. Finally, for each k and each window size a single misclassification error per image (SCE) as well as the mean misclassification error over all images (MCE) is computed to evaluate the results of our approach. The SCE is estimated by
84 König et al. the ratio of the number of correctly classified pixels to the number of falsely classified pixels of the ROI. The mean misclassification error is then computed by averaging of the SCE s. 3 Results Fig. 3 shows the MCE for all window sizes for k = 64 (MCE#k64) as well as the minimum (MCE#Min) and maximum mean misclassification error (MCE#Max) for each window size for all k-values. Our approach reaches its lowest error rate of 7.80% at a window size of 180 px 180 px used for feature extraction and training of the classifier. The smallest window size of 80 px 80 px produces the highest error rate with 25.16 %. The MCE differs with respect to the parameter k only between 0.38 % and 3.08 % depending on particular window sizes. The results of the SCE for a fixed window size of 180 px 180 px and a fixed k = 64 are shown in Fig. 4. The results can be used for classification into pathological and healthy tissue by assigning the class of the majority of pixels in the ROI. Most (60 of 63 cases) of the evaluated images reach a low misclassification error 30 %, thus, more than 70 % of the ROI pixels are classified correctly. In 29 of 63 cases more than 99 % of the pixels were classified correctly. Only 3 of 63 ultrasound images (4.76 %) show a misclassification rate 50 %. 4 Discussion Using non-directional wavelet transform features and first-order grey-level intensity statistics combined with the presented approach of computing features for 0.3 MCE#Min MCE#Max MCE#k64 0.25 0.2 MCE 0.15 0.1 0.05 0 80 100 120 140 160 180 200 220 Window size Fig. 3. Results of the MCE for all analyzed window sizes for k = 64 (MCE#k64). The upper red line displays the maximum mean misclassification error (MCE#Max) of all k-values, the lower green line shows the results of the minimum mean misclassification error (MCE#Min) of all k-values.
Detection of Myositis in Ultrasonographies 85 each pixel of the ROI with various window sizes provides a stable texture-based analysis for detecting myositis in ultrasonographies. The choice of the window size impacts the classification results. Too small windows (range from 80 px 80 px to 150 px 150 px) result in worse classification rates (Fig. 3) since those windows do contain enough relevant information. Too large windows (range from 190 px 190 px to 220 px 220 px) carry the risk that too many other kinds of tissues and structures are observed within the selected window and thus the results can be falsified strongly. Determining the proper window position and size also depends on the location of the muscle tissue in the image, on the resolution of the image and on the kind of scanned muscles. There might be no optimal solution for all data sets. Hence, some imaging protocol should be used for image generation that fixes important parameters such as the anatomic region so that different windows sizes can be determined for different protocols. The selection of the k-values of knn has little influence on the classification rate. However, smaller k-values result in slightly worse classification rates than higher k-values. As mentioned, a window size of 180 px 180 px resulted in best classification rates. Nevertheless, 3 of 63 images have been misclassified (Fig. 4) because those images include feature characteristics that are inconsistent with our initial assumptions that the texture of pathological muscle tissue images is more diffuse with smaller contrast changes compared to the texture of healthy muscle tissue images with structured tissue and higher contrast changes (Fig. 5). A reason for this could be the different procedures which were used during image acquisition or the fact that our data set contains images from different muscles. Although our data set is insufficient for clinical reliable results, we have shown that texture analysis can be used as a robust, automatic method to anal- 1 0.9 SCE#H SCE#P 0.8 0.7 0.6 SCE 0.5 0.4 0.3 0.2 0.1 0 0 10 20 30 40 50 60 Classification of images Fig. 4. Results of the SCE for a fixed window size of 180 px 180 px and a fixed k = 64. The green bars represent the SCE of healthy muscle tissue images (SCE#H) and the red bars of the pathological muscle tissue images (SCE#P).
86 König et al. Fig. 5. Examples of misclassified images for healthy (left, SCE of 81.30%) and pathological (right, SCE of 62.88%) muscles. yse myositis from ultrasound images. Future work will focus on an evaluation of a larger number of images including more patients acquired with consistent imaging protocols to prevent variations caused by different acquisition parameters, positions, and orientations of the transducer or anatomical regions as well as maximization of unpreventable variations within the data like patient specific variation, i.e. age, gender and fat rate, to get a representative sample set. Furthermore, an application of the proposed method to other muscles could be possible, if one would select different window positions and sizes for each muscle group. An integration of staging of myositis is a topic which is worth of further research. Different classification techniques will be evaluated, i.e. SVM, as we decided on the knn classifier because of visual separability of our feature space. However, this assumption could be a misinterpretation of the visual observations caused by the dimensional down-projection of the feature space for visualization. References 1. Pillen S, van Alfen N. Skeletal muscle ultrasound. Neurol Res. 2011;33(10):1016 24. 2. Wu CM, Chen YC, Hsieh KS. Texture features for classification of ultrasonic liver images. IEEE Trans Med Imaging. 1992;11(2):141 52. 3. Huang YL, Wang KL, Chen DR. Diagnosis of breast tumors with ultrasonic texture analysis using support vector machines. Neural Comput Appl. 2006;15(2):164 9. 4. Tsiaparas NN, Golemati S, Andreadis I, et al. Comparison of multiresolution features for texture classification of carotid atherosclerosis from B-mode ultrasound. IEEE Trans Inf Technol Biomed. 2011;15(1):130 7. 5. Pohle R, Fischer D, von Rohden L. Computergestützte Gewebedifferenzierung bei der Skelettmuskelsonographie. Ultraschall Med. 2000;21(6):245 52. 6. Mallat SG. A theory for multiresolution signal decomposition: the wavelet representation. IEEE Trans Pattern Anal Mach Intell. 1989;11(7):674 93. 7. Cover T, Hart P. Nearest neighbor pattern classification. IEEE Trans Inf Theory. 1967;13(1):21 7.