Abstract:
In this research we have developed and evaluated a computer aided diagnosis (CAD) model that is based on the automated segmentation of breast lesion on ultrasound images. Active contour model has been used for segmentation after removal of speckle noise followed by morphological and textural features computation. The CAD model is based on the breast imaging reporting and Data system (BI-RAD) for major feature selection to classify benign-malignant group and benign-malignant-normal group. Ultrasound images were collected from multiple hospitals with their consent. The data set consisted of 163 actual ultrasound images of benign, malignant and normal images used for the analysis. In the proposed method, after segmentation and feature extraction test image is placed into either of two groups of benign-malignant group and benign-malignant-normal group. Binary support vector machine classifier has been used to lesion based identification of breast tumor as benign or malignant in the first group, whereas multiclass SVM using one vs. all method has been used for similar identification in the second group. All cases were samples with k-fold cross validation method, performance of the classifier for classifying were evaluated by receiver operating characteristics in both groups. The area under the curve for benign-malignant group using morphological features were 0.97, classifier accuracy to be 94% with sensitivity and specificity of 97% and 88%.In benign-malignant and normal group areas under the curve for benign, malignant and normal group was 0.94%, 0.84 and 0.86 % with a sensitivity and specificity of 94% and 83%. It was concluded that proposed CAD model was able to differentiate, with acceptable accuracy, benign from malignant breast tumor using morphological features and normal from benign and malignant using textural features. Hence this model can be used to detect tumor and can provide a reliable second opinion to the radiologist. Key words: breast imaging reporting and Data system (BI-RAD),
