Multi-Centric Comparison of Deep Learning Models for Lesion Detection in Breast MRI

Abstract

Breast magnetic resonance imaging (MRI) is a common modality for diagnostic imaging in breast cancer, creating a need for automated image analysis to assist in early detection and diagnosis. In this study, we compare multiple deep learning-based segmentation and detection algorithms for lesion detection in dynamic contrast-enhanced (DCE) breast MRI. We utilized a large multi-centric dataset comprising T1-weighted DCE MR images from nine clinical sites across seven countries, encompassing diverse imaging characteristics and scanner types. We evaluated several models, including the standard nnU-Net, an adapted nnU-Net with modifications to reduce false positives, a coarse-resolution version thereof, the transformer-based SwinUNETR-V2, and nnDetection. The standard nnU-Net achieved a high lesion-level sensitivity of 83.8% but produced an average of 3.334 false positives per case, which is impractical for clinical use. The adapted (coarse) nnU-Net significantly reduced false positives to 0.666 (0.397) per case with a slight decrease in sensitivity to 79.9% (75.8%). SwinUNETR-V2 achieved comparable performance to the adapted nnU-Net. nnDetection outperformed nnU-Net in the high-sensitivity region, but performed worse than the adapted models in the lower-sensitivity region, with respect to false positives. To conclude, the nnU-Net again provides a good baseline, but our lesion detection task motivates adaptations to reduce the number of false positives.