Classifying region of interests from mammograms with breast cancer into BIRADS using Artificial Neural Networks

Abstract

Breast cancer is one of the most common cancers among female diseases all over the world. Early diagnosis and treatment is particularly important in reducing the mortality rate. This research is focused on the prevention of breast cancer, therefore it is important to detect micro-calcifications (MCs) which are a sign of early stage breast cancer. Micro-calcifications are tiny deposits of calcium which are visible on mammograms as they present as tiny white spots. A computer-aided diagnosis system (CAD) is created with the development of computer technology that way radiologists are aided improving their diagnostics while using CAD as a second reader. We are aiming to classify into BIRADS 2, 3 and 4 which are the stages when the cancer can be prevented and a fourth category called No lesion which are veins and tissue that our high pass Gaussian filter detects. This research focuses on classification using ANN (Artificial Neural Network). Experimenting with the categories to classify into using ANN, the results were the following: Into the four mentioned before an overall accuracy of 71% was obtained, then joining categories BIRADS 2 and 3 into one and classifying into 3 categories gave an 80% of accuracy. Joining this two categories was the result of analizing the ROC curve and observation of the ROI images of the MCs as the regions measured are very alike in this two categories and variation is that MCs are more present in BIRADS 3 than in BIRADS 2. Data matrix was reduced using PCA (Principal Component Analysis) but it did not gave better results so it was discarded as the ANN accuracy to classify was reduced to a 69.8%.