Fractal feature based early breast abnormality prediction

Abstract

Breast cancer is associated with high mortality rates in women of both developing and under developed countries. Moreover, due to the poor medical facilities and lack of awareness, this mortality rate is higher in rural areas than that of the urban areas. Hence, to reduce this high mortality rate, the early detection of the breast diseases before the onset of the cancerous mass is very crucial. Among various breast imaging modalities, X-ray Mammography stands out to be the gold standard modality for Breast cancer detection. But vulnerability of women below 40 years towards radioactive exposure of X-ray mammography necessitates the concerned research community to explore avenues devoid of radioactive hazard as well as preferably non-invasive. Infrared Thermography (IRT) meeting such important requirements can be used as an adjunctive tool in breast abnormality detection of women of all age groups. Besides, due to its portability and cost effective nature, it can be used as a routine checkup tool for patients in remote areas and thus, can point out the subjects who requires urgent medical attention. To validate the predictability of both mammography and thermography in breast cancer detection, this paper develops a suspicious region based breast abnormality detection system. The paper investigates the efficacy of fractal features over the most widely used texture features in anomalous region based breast abnormality prediction from both mammograms and thermograms. We focus on fractal features in discriminating the abnormal and severe abnormal breast images from the normal and mild abnormal breast images by observing the difference in fractal dimension and lacunarity values. We investigated that the combination of fractal dimension and lacunarity features gives prediction accuracy of 95.94% on the mini-MIAS mammogram dataset of 128 images and 86.11% on a newly created DBT-TU-JU breast thermogram dataset of 36 abnormal images as compared to 79.31% and 78.94% using texture features, respectively. The experimental results reveal that the fractal features are more efficient in disease affected region based breast abnormality prediction from both mammograms and thermograms.