Comparison between HOG and Haar descriptors in the detection of abnormal tissue in mammograms

Abstract

The design and development of artificial intelligence and machine learning models applied to medical images are an alternative for the detection and classification of abnormal clinical patterns. Mammography images help identify abnormal areas of dense breast tissue that indicate risk factors for breast cancer. In this article, we compare the HOG and Haar descriptors by varying the negative sample factor parameter in training a machine learning based cascade object detector using MATLAB®. The objective was to identify the best descriptor and parameters that allow increasing the precision of detection and labeling of regions of clinical interest due to the presence of abnormal regions in digital mammograms. The images used in the training were obtained from the free database of the United Kingdom Mammography Image Analysis Society (MIAS) and tests were performed with images of Breast Cancer Digital Repository (BCDR). The HOG and Haar descriptors were used in 15 stages, with a different value in the negative sample factor parameter in each descriptor. The results showed that using the HOG descriptor with a low value of negative sample factor, the precision in detecting abnormal tissue in mammography was higher compared to the use of Haar descriptor.