Computer-Aided Diagnosis of Breast Tumors Using Textures from Intensity Transformed Sonographic Images

Abstract

The malignancy of breast tumors are evaluated via ultrasound images on clinical examination. As a second viewer, a computer-aided diagnosis (CAD) system was developed to classify the breast tumors using texture features to avoid misclassifying carcinomas. A total of 69 cases including 21 malignant and 48 benign masses were acquired. For intensity-invariant texture extraction, the ultrasound images were first transformed into ranklet images to reduce the effect of brightness variability. From the ranklet images, tumor texture and speckle texture were extracted and compared to those from the original ultrasound images for tumor diagnosis. In the trade-offs between sensitivity and specificity, the ranklet-based tumor texture and speckle texture were all significantly better than those of the original US images (Az: 0.83 vs. 0.58, p-value=0.0009 and Az=0.80 vs. 0.56, p-value=0.02). The proposed CAD system using textures from intensity transformed sonographic images is robust to various gray-scale distributions and is more suitable in clinical use.