Statistical Analysis for Radiologists’ Interpretations Variability in Mammograms

Abstract

Conventional mammography is considered the modality of choice for the detection of breast cancer. The process involves a human radiologist visually diagnosing the mammogram, which causes limitations such as missing a cancer and/or diagnosing a false cancer. Another disadvantage of conventional mammography is the variability among screening radiologists in interpreting mammographic images. The objectives of this study are to verify this variability and to develop an image processing algorithm that can automatically detect benign tumors of the female breast. A sample of ten digital mammograms obtained from the MiniMIAS database was distributed to four different radiologists in order to verify the variability among them. Furthermore, three algorithms were developed in order to automatically detect benign tumors of the female breast. The proposed algorithms were based on combinations of certain statistical features and were tested on the same sample of images. Results showed that the detection mechanism using the proposed algorithms was acceptable despite the fact that they exhibited a few errors. It was concluded that the use of a combination of the mean and median statistical tools is effective in assisting radiologists in interpreting mammographic images containing benign tumors.