Research On The Pre-processing For The Breast Cancer Automatic Diagnosis System

Breast cancer is a common form of malignant tumors in women. It shows clearly that early discovery, early diagnosis and early treatment of breast cancer can significantly increase the chance of survival and reduce breast cancer mortality for patients. Along with the fast development of computer imaging technology, the computer aided system for finding microcalcification in digital mammograms has become the key techniques in the field of early diagnosis of breast cancer. Microcalcifications are early sign of breast cancer appearing as isolated bright spots in mammogram images. However, they are difficult to be detected due to their small sizes, noisy and big image background. Identifying regions of interest (ROI) could shrink the areas for detecting microcalcifications or even screen out totally normal images without microcalcifications in it. Because in the all-pervading breast cancer examination, images of healthy people account for 99.5%. For that reasons, this work is very meaningful for computer aided diagnosis system treating with mammograms. In this paper, we propose a novel method for identifying ROIs for microcalcifications. At first, we use simple gray level distribution and cost-sensitive ANN to segment out the background and smooth tissue. Then, we enhance the remained part of the image with Morphological bandpass filter(MBF) and Laplacian-of-a- Gaussian(LoG) filter. At last, a thresholding method is used to get the areas bearing microcalcification clusters. Our method can obtain accurate ROIs. We carry out a series of the experiments on 208 images from the breast cancer database of Nanjing Zhongda Hospital. The experimental results confirm that with our method, an average consuming time of 101s and an average discarding rate of 99% can be achieved. We also screen out 24 images from the 70 normal images without microcalcifications and obtain the average normal image discarding rate of 34%.