Similar Intensity-Based Euclidean Distance Feature Vector for Mammogram Image Classification

Mammogram imaging is economical, easily available, non-invasive and preferred for breast cancer detection. This paper proposes and tests a new feature vector based on the sum of Euclidean distances of similar subsequent pixels on various well-known classifiers. The preprocessed and augmented region of interest images extracted from the Mammographic Image Analysis Society (MIAS) dataset are used for performance evaluation. The classifier’s input data is balanced by randomly selecting 2000 images from each normal, benign and malignant categories. On an Ensemble Subspace KNN classifier using a tenfold cross-validation approach, it achieved good sensitivity/recall, specificity, precision and F1-scores along with a classification accuracy of 98.4%. This technique can be used for regular breast screening for early-stage breast cancer detection and breast abnormalities identification. It can prioritize mammograms for further analysis as well as provides an opinion to the radiologists in decision-making.