Enhancing Deep Learning-Based Breast Cancer Classification in Mammograms: A Multi-Convolutional Neural Network with Feature Concatenation, and an Applied Comparison of Best-Worst Multi-Attribute Decision-Making and Mutual Information Feature Selectio

Breast cancer remains a leading cause of mortality among women worldwide wherein there is a need for innovative diagnostics to enhance early detection accuracy. The primary goal of this research is to improve the accuracy and other performance metrics of breast cancer classification in mammograms with existing deep-learning techniques. This study proposes an enhanced deep learning-based approach for breast cancer classification in mammogram images, leveraging multiple Convolutional Neural Networks (CNNs), Best-Worst Multi-Attribute Decision-Making (BWM-MCDM) feature selection, and feature concatenation. Eight pretrained CNN models (InceptionV3, ResNet50, MobileNet, VGG16, VGG19, Xception, DenseNet169, and EfficientNetB7) were fine-tuned on an augmented Mammographic Image Analysis Society (MIAS) dataset. Features for each image in each trained model were extracted and those features with zero variance were eliminated. These extracted features underwent two feature selection methods, BWM-MCDM and Mutual Information (MI) feature selection. All features from BWM-MCDM were concatenated, features from MI were concatenated, and the naïve (all) features—without feature selection but with zero variance removed—were also concatenated. A total of three sets of concatenated features were gathered. These three concatenated features were each used to train a neural network classifier. Results for the proposed framework with BWM-MCDM demonstrated a reduction in feature dimensionality with the deemed ability to retain critical information. From five cross-folds, the mean for each performance metric achieved an accuracy of 98.74% ($\mathrm{SD}=0.0020$), F1-score of 98.73% (S D=0.0021), ROC-AUC of 99.80% ($\mathrm{SD}=0.0004$), and MCC of 95.31% ($\mathrm{SD}=0.0077$), all surpassing the eight individual CNN models’ performance. A comparative analysis using the Friedman and Nemenyi post-hoc tests revealed that MI consistently outperformed BWM-MCDM, although both performed similarly compared to naïve features, implying that feature concatenation is effective as itself. However, concerning the ratio of the number of retained extracted features and the performance metrics, BWM-MCDM outperforms both MI and naïve features. This just shows that factoring in the number of extracted features by each feature selection methods, respectively, changes the overall picture in the overall performance ratings based on the different performance metrics. The findings highlight the effectiveness of combining multi-CNN feature extraction, BWM-MCDM feature selection, and feature concatenation for breast cancer classification. This innovative approach has the potential to significantly improve early detection and overall predictive accuracy, contributing to advancements in breast cancer diagnosis and treatment.