An Integrated Approach for Improved Mammography Image Classification Using EfficientNet and Attention Mechanism

In this study, a deep learning method utilizing the EfficientNet-BO model has been developed for the purpose of classifying mammography images into three categories: benign, malignant, and benign without callback. This approach is designed to address current challenges in mammography-based diagnostics, such as the limited amount of available data for training and the difficulty of distinguishing subtle lesion features in dense breast tissue. The model has been augmented with Test-Time Augmentation and an attention mechanism that jointly highlights lesion features with the masked image, thereby increasing the focus on key diagnostic regions. Experimental results on the CBIS-DDSM dataset demonstrate that the proposed model effectively overcomes the challenges associated with limited data and small features, resulting in a substantially improved classification performance over the base model with 96.5% accuracy.