Grand Challenge: Digital Breast Tomosynthesis Lesion Detection

Abstract

Our team at NYU developed an object detection system based on EfficientDet, leveraging a large dataset of 2D mammograms and DBT volumes to pre-train our models. We used an ensemble of 35 models to enhance robustness and improve lesion detection performance. The competition highlighted the effectiveness of 2D slice-based detection models, the importance of additional training data, and the potential of ensemble methods in medical imaging AI.

Jungkyu Park

PhD Candidate