Reproduce “From 2D to 3D Without Extra Baggage: Data-Efficient Cancer Detection in Digital Breast Tomosynthesis” on OpenResearch

About this paper

Yen Nhi Truong Vu, Dan Guo, Sripad Joshi, Harshit Kumar, Jason Su, Thomas Paul Matthews

Whiterabbit.ai's M&M-3D introduces a data-efficient deep learning architecture for 3D breast cancer detection in Digital Breast Tomosynthesis (DBT) by extending a 2D mammography model without adding new learnable parameters. This method enables sophisticated 3D reasoning while leveraging 2D pretraining to achieve state-of-the-art performance on public benchmarks, even with limited annotated DBT data.

Key insights

• Learnable and sophisticated 3D reasoning can be achieved in deep learning models without introducing new, learnable parameters, allowing effective transfer learning from existing 2D-pretrained models.
• Dynamic, malignancy-guided feature fusion across 2D slices is crucial for building robust 3D representations that reduce false positives and provide accurate, interpretable z-axis localization.
• Leveraging abundant 2D Full-Field Digital Mammography (FFDM) pretraining is paramount for achieving high performance and data efficiency in 3D DBT cancer detection, especially in low-data and low-annotation regimes.

Summary by alphaXiv.