Transformer-based Transfer Learning for Breast Cancer Detection via Digital Breast Tomosynthesis
Mikea Dimech · Zenodo (CERN European Organization for Nuclear Research) · 2023
Breast cancer has emerged as the most prevalent form of cancer worldwide and the leading cause of mortality from any cancer in women, emphasising the critical need for its early detection. Traditional 2D mammography has long been the cornerstone of breast cancer screening, given its proven efficacy in early prognosis and reduction of mortality. However, the limitations of this modality have precipitated the development and adoption of more advanced imaging techniques, such as Digital Breast Tomosynthesis (DBT), which yields a 3D view of the breast. The application of deep learning to classify these scans is paramount to increasing the efficiency and efficacy of screening. However, this task presents arduous challenges such as high computational demands, scarcity of diverse data, and generalisation across varying demographics and imaging equipment. Transfer learning has proven to be effective in overcoming these challenges by enabling models to leverage pre-existing knowledge and adapt to new tasks. However, its application to DBT has been confined to Convolutional Neural Network (CNN) architectures. Meanwhile, the Transformer paradigm, which has revolutionised various deep learning domains and shown promise in traditional mammography, has yet to be explored in the context of DBT via transfer learning, highlighting a notable research gap. This study thus investigates the performance of a pre-trained Transformer model, fine-tuned on a public DBT dataset, comparing it with a fine-tuned state-of-the-art CNN. While the Transformer outperforms the convolutional architecture, the results reveal severe challenges posed by the imbalanced dataset, with models developing biases towards the majority class. Data augmentation, while beneficial, could not fully address the issue, and the need for more diverse datasets became clear. Moreover, the results underscore the importance of preserving the high resolution of DBT images, with fine-grained details, such as subtle changes in breast tissue captured at higher resolutions, proving to be pivotal for accurate diagnosis. Finally, a comparative analysis between CNNs and transformer-based models revealed that Transformers excel in recognising broader patterns across the entire image, which aids to mitigate overfitting. By using pre-trained transformers and fine-tuning them on this dataset, their effectiveness was highlighted, suggesting they can match or even surpass CNNs in breast cancer detection via DBT.