Enhancement of Breast Mammography to Rapid Screen Abnormalities Using 2D Spatial Fractional‐Order Feature Extraction and Multilayer Machine Vision Classifier

Abstract

While breast tumor symptoms [benign (B) or malignant (M) tumors] or lymph node metastasis can be detected at an early stage, the timely discovery of abnormalities could simultaneously improve survival rates and enhance therapeutic efficacy. During first‐line mammography examination, upon finding any solid‐white region, clinicians or radiologists can manually select the region of interest (ROI) from lateral‐view images with a specific bounding box (BB). This study used two‐dimensional (2D) fractional‐order convolution (FOC) operations with fractional‐order parameters (v = 0.3−0.4) to sharpen and retain the structure of possible focus, inhibit the features of background tissue, and remove unwanted noise. The gray‐scale values of breast mammography can be readjusted to enhance the image contrast using a nonlinear intensity transformation function. Abnormal contour levels can automatically be searched on the basis of the edge detection of normalized gradients to interpret the main ROI region using a contour algorithm. Hence, the main ROI can be rapidly identified on each breast mammography. The central region of the main ROI can then be divided into five sub ROIs, which are fed to a multilayer machine vision classifier. We used a general regression neural network (GRNN)‐based classifier to separate normal (no tumor) images from abnormal (B or M tumor) ones and obtain an automatic screening support tool that can decisively confirm suspicious lesions and achieve accurate detection. Using images from the Mammographic Image Analysis Society (MIAS) digital breast mammogram database, we applied the proposed GRNN‐based classifier with K‐fold cross‐validation to emphasize the possible focus and provide additional confidence during imaging examination. Results showed an average recall (%), an average precision (%), an average accuracy (%), and an average F1 score of 87.31%, 90.50%, 88.67%, and 0.8883%, respectively, for the rapid screening of B and M tumors on digital breast X‐ray images. The GRNN‐based classifier is superior to traditional techniques in terms of adaptive learning schemes, computational time consumption, and design cycle. © 2021 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.