Predicting the BI-RADS Lexicon for Mammographie Masses Using Hybrid Neural Models
Juanita Hernández-López, Wilfrido Gómez‐Flores · 2020
The mammographic image analysis carried out by radiologists describes findings to infer the type of breast pathology to lead patients to adequate treatments. For such an analysis, the experts use a lexicon in the Breast Imaging-Reporting and Data System (BI-RADS), through which the shape and margin of the tumor are described, as well as the breast parenchymal density to determine the risk of developing breast cancer. In this sense, the development of technological tools to assist radiologists in the mammographic description is justified. In this article, two hybrid classification techniques based on neural models are proposed to predict the BI-RADS lexicon attributes for masses. These hybrid models comprise a pre-trained convolutional neural network (AlexNet) for feature extraction and neural models for pattern classification, namely, multilayer perceptron (MLP) and dendrite morphological neuron (DMN). The obtained classification results suggest that the hybrid neural model schemes obtain an accuracy of 91% for predicting the shape and density attributes and an 85% accuracy for predicting margin attributes. The Wilcoxon test indicates no significant differences between both hybrid models. Therefore, the proposed models can be useful for predicting attributes included in the BI-RADS lexicon for masses.