Improving Mammography Lesion Classification Based on Transfer Learning with EfficientNet-B0

Breast cancer remains a critical public health concern, affecting approximately one in eight women in the United States [21]. This paper presents a historical overview of advancements in breast cancer diagnosis, with a particular focus on the evolution of radiological studies and imaging technologies. To address the scarcity of medical imaging data, this study employs transfer learning (TL) in conjunction with the EfficientNet-B0 model to classify mammographic images. The proposed framework, termed Mammography Lesion Transfer Learning Classification (MLTLC), leverages a robust deep learning pipeline for improved diagnostic accuracy. The Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CIS-DDSM) serves as the primary dataset, with extensive data preprocessing and augmentation applied to enhance model compatibility. This paper provides a comprehensive discussion of the dataset, emphasizing inherent challenges associated with medical image processing. The CBIS-DDSM dataset comprises mass and calcification images, with 1,231 mass samples utilized for training the pre-trained EfficientNet-B0 model, while 1,227 calcification images from the training set for the TL-based classification pipeline. The proposed model achieves an average accuracy of 95.59% on the mammographic calcification test set, demonstrating its efficacy in breast cancer classification. The results highlight the model's capability to precisely differentiate between malignant, benign, and benign without callback cases, underscoring its potential for clinical applications in early breast cancer detection.