부산대학교 도서관

Breast cancer is a prevalent form of cancer with significant mortality and morbidity rates among women worldwide. Early detection is vital in increasing the chances of survival and one of the major approaches for breast cancer screening and detection is medical imaging. Advancements in technology have given rise to 3D medical imaging such as Abbreviated Breast MRI and DBT, overcoming the challenges of tissue overlapping in 2D modalities. However, deep learning model development using this 3D medical imaging comes with higher computational costs and complexity. This study proposes a lightweight deep learning technique based on three slice selection techniques using entropy, variance, and gradient magnitude values from DBT medical imaging modality. The selection techniques help select only the most informative slices making computational cost and complexity reduced. Entropy value-based slice selection performed best with an accuracy of 91%. The results obtained using the slice selection techniques for lightweight deep learning model development show that it can diagnose breast cancer with a lower number of slices and less computational complexity compared to existing methods.