Hybrid mammogram mass segmentation and classification using efficient handcrafted features and ensemble learning

Abstract

Early detection and accurate classification of mammographic abnormalities are essential to reducing the mortality rate. The proposed work on mammogram mass segmentation addresses this issue using a hybrid approach that combines mean median iterative filtering, fuzzy logic, Multilevel thresholding, and morphological operations. Handcrafted feature extraction techniques, including pass band Discrete Cosine Transform (PDCT), Wavelet Gray Level Co-occurrence Matrix (WGLCM), and Multiscale Local Binary Patterns (MLBP) are applied to extract texture, shape, and intensity features of the segmented mass. Furthermore, the Chi-square statistic is utilized for feature selection from the extracted features. An ensemble classifier is used to identify the class for mass features, that is, a combined class based on the majority class obtained from Linear SVM, Decision Trees (DT), Random Forest, Naive Bayes, and K-Nearest Neighbors (KNN). Extensive experiments are conducted using the MIAS and DDSM mammogram benchmark datasets to test the performance of the proposed method. Cropped Suspicious regions from full mammograms are arranged based on four types of Density (D, E, F, and G) and BIRADS classes (as a normal–abnormal case and then the abnormal case further as a benign or malignant class). The results for the PDCT method applied to the MIAS Dataset demonstrate high accuracy, specificity, and sensitivity in distinguishing between normal and abnormal masses, as well as benign and malignant masses. The findings indicate that PDCT features significantly outperformed the classification tasks for both normal–abnormal and benign–malignant categories. The major contributions of this work include integration of complementary handcrafted classifier an effective feature selection strategy and an ensemble classification framework that improve mammographic mass classification performance. This highlights the effectiveness of the PDCT method in improving breast cancer diagnosis through precise feature extraction and classification.