Enhancing Breast Cancer Mammography Categorization using Two Methods: A Deep Learning Structure Combining Fully Connected Layers, SMOTE, and ResNet50 for Balanced and Unbalanced Data

Abstract

Breast cancer remains a major global health issue, where early and precise diagnosis plays a critical role in improving patient outcomes. Mammogram imaging is widely used for detection, but interpreting these images requires expert knowledge and can be both time-consuming and complex. Although deep learning techniques have shown strong potential in medical image analysis, one of the key challenges is the presence of imbalanced datasets, which often leads to biased and less reliable models.
In this work, we propose a novel deep learning framework for breast cancer classification using mammogram images. The proposed system introduces a dual-module strategy to effectively address data imbalance using the Synthetic Minority Over-sampling Technique (SMOTE). In the first module, SMOTE is applied to the entire dataset to create a balanced training set. In the second module, 20% of the original imbalanced data is reserved for evaluation, while SMOTE is applied to the remaining 80% to improve model learning.
The framework combines a block-wise Convolutional Neural Network (CNN) architecture, where VGG16 is used for input preprocessing and standardization, and ResNet50 is employed for robust feature extraction. The extracted features are then passed through a fully connected classification network consisting of multiple dense layers, along with batch normalization and dropout techniques to reduce overfitting. After several iterations, the final model includes three dense layers with 128, 256, and 128 neurons, and a dropout rate of 0.5.
Experimental results demonstrate that the proposed model achieves an accuracy of 99% on the balanced dataset and 90% on the imbalanced evaluation set. Additionally, the framework incorporates an interpretable visualization mechanism to analyze predictions across different classes, improving transparency in decision-making.
Overall, this approach significantly enhances the accuracy and reliability of breast cancer detection from mammogram images. By effectively handling data imbalance and providing interpretable outputs, the proposed framework contributes to the advancement of computer-aided diagnosis systems and can be extended to other medical imaging applications