MRI Tumor Segmentation applying Different Wavelet Transform Features, Sparse Representation-Based Classifier and Snake Algorithm
Nooshin Nabizadeh (University of Miami)
Characterization and Imaging of Structural and Material Imperfections
Mon 4:20 - 5:40
Barus-Holley 191
In this paper, brain MR images segmentation of tumors using an efficient multi-stage method is proposed. This method employs only one MR modality as Tl-weighted sequences. The brain midline and histogram asymmetry between two brain hemispheres is used to detect the tumor slice. From each selected slice, four rectangular windows are extracted from each hemisphere, which covers brain tissues. The location of windows change based on the size of the brain in each slice. Applying Discrete Cosine Transform (DCT), Discrete Wavelet Packet Transform (DWPT), Dual Tree Complex Wavelet Transform (DTCWT), and Complex Morlet Wavelet Transform (CMWT), features are extracted from each window. DTCWT is designed to overcome some shortcomings of DWT in image processing such as oscillations around singularity, shift variance, aliasing after processing on wavelet coefficients, and lack of directionality. The advantage of DTCWT in texture analysis is it is not sensitive to location of texture patterns. It also can capture more directions of texture patterns, which is efficient in lesion segmentation. The lowpass sub band and the highpass subband for each level, and the lowpass coefficients at every scale are calculated as features. CMWT has a form very similar to the Gabor-wavelet transform. The important difference is that the window function dose also be scaled by the scaling parameter, while the size of window in Gabor transform is fixed. Energy of each complex Morlet kernel at different scales and orientations are calculated. Feature selection on 4102-dimensional feature vector is performed using principle component analysis. The selected features are then classified by using Sparse Representation-Based Classifier (SRC). The windows which are classified as tumor windows are employed for initial tumor segmentation. The final tumor segmentation is accomplished by the snake algorithm. The method is applied over hundreds of Tl-weighted sequences and the results are acceptable.