Accurate Automated Breast Cancer Detection and Classification Using a Hybrid Deep Learning Model

Abstract

In order to classify breast cancer, the existing models used are computer diagnostic approaches, statistical and rule-based models, traditional machine learning models, standalone deep learning models, and segmentation-based models. Among these, traditional rule-based suffers in poor generalizability and require manual tuning. Traditional machine learning approaches depend more on handcrafted features and have limited support for heterogeneous datasets of imaging. The independent deep learning models, which are computationally intensive using CNN and prone to overfitting due to limited data. Segmentation methods consume more annotation costs, which increases the system complexity. To overcome these, a hybrid deep learning framework is needed that uses complementary features learning mechanisms to process both low-level and high-level discriminative characteristics. The involvement of multi-scale CNN captures in parallel both fine-grained texture features and coarse structural patterns, for improved discrimination. The additional adaptive attention mechanism highlights clinically relevant regions by background noise suppression, enhances precision and sensitivity, and minimizes false alarms. Data preprocessing, including noise reduction, contrast enhancement, and data augmentation, is used to enhance generalization capability. For dataset imbalance handling, synthetic minority oversampling and elastic augmentation are preferred. For optimization, label smoothing and dropout scheduling of regular aware training are used. The proposed approach works with multi-level feature extraction, optimal training, and robust learning, making the model more effective than other models considered.