Abstract

Breast cancer is the most common of all cancers and second most deadly cancer in women in the developed countries. Mammography and ultrasound imaging are the standard techniques used in cancer screening. Mammography is widely used as the primary tool for cancer screening, however it is invasive technique due to radiation used.

Ultrasound seems to be good at picking up many cancers missed by mammography. In addition, ultrasound is non-invasive as no radiation is used, portable and versatile. However, ultrasound images have usually poor quality because of multiplicative speckle noise that results in artifacts. Because of noise segmentation of suspected areas in ultrasound images is a challenging task that remains an open problem despite many years of research.

In this research, a new method for automatic detection of suspected breast cancer lesions using ultrasound is proposed. In this fully automated method, new de-noising and segmentation techniques are introduced and high accuracy classifier using combination of morphological and textural features is used.

We use a combination of fuzzy logic and compounding to denoise ultrasound images and reduce shadows. We introduced a new method to identify the seed points and then use region growing method to perform segmentation. For preliminary classification we use three classifiers (ANN, AdaBoost, FSVM) and then we use a majority voting to get the final result. We demonstrate that our automated system performs better than the other state-of-the-art systems. On our database containing ultrasound images for 80 patients we reached accuracy of 98.75% versus ABUS method with 88.75% accuracy and Hybrid Filtering method with 92.50% accuracy.

Future work would involve a larger dataset of ultrasound images and we will extend our system to handle colour ultrasound images. We will also study the impact of larger number of texture and morphological features as well as weighting scheme on performance of our classifier. We will also develop an automated method to identify the "wall thickness" of a mass in breast ultrasound images. Presently the wall thickness is extracted manually with the help of a physician.