Rule Based Fuzzy Image Segmentation for the Detection of Breast Cancer from Ultrasound Image

Early detection of breast cancer is the most important to reduce the number of deaths among women. Computer aided diagnosis plays a vital role in all clinical diagnosis and hence used in the proposed work for detection of breast cancer. To reduce the speckle noise in ultrasound image Median filter, Non Local Means filter and Lee filter was applied for preprocessing. The non-Local means filter had been used as it provides the highest PSNR values. Fuzzy clustering method is applied for the segmentation of the denoised image. After segmenting the image into set of clusters fuzzy level set algorithm is applied for more accurate detection of edges in the tumour region. PSNR value of 35.86dB had been obtained after denoising using Non Local mean filter. The mean, entropy and standard deviation parameters are analyzed for the different cluster size of the benign and malignant image. From the results it had been observed that the cluster size 4 provides better segmentation as it provides almost constant parameters for different images. From the cluster that belongs to the region of interest, fuzzy level set algorithm had been applied for minute edge detection. The segmented image after applying fuzzy level set provides better perception compared to the image without level set. After the segmentation, in the feature extraction, important features such as edge, intensity, contrast and orientation are extracted using Feature-based morphometry approach (FBM). Specifically to extract orientation, the images are scaled at 0o, 45 o, 90 o and 135 o using Gabour filter. The features such as mean, standard deviation and entropy are calculated for all the seven features and the results are compared for more number of benign and malignant images. These extracted features are used for the classification stage. In the classification, 50 ultrasound breast cancer images consist of 14 benign images and 36 malignant images are used. The images are trained by Support Vector Machine using the Generalized Multiple Kernel Learning with the help of regularization 0 and 1. From this training, the maximum accuracy, sensitivity, specificity and BAC obtained as 73, 100, 38 and 69 respectively with regularization 1.