Multi-institutional classification of lesions in breast MRI: a comparison of one-stage and two-stage methods.
Authors
Affiliations (14)
Affiliations (14)
- Fraunhofer Institute for Digital Medicine MEVIS, Bremen, Germany.
- Siemens Healthineers AG, Research & Clinical Translation, Magnetic Resonance, Erlangen, Germany.
- Siemens Healthineers AG, Digital & Automation, Forchheim, Germany.
- Diagnostic Breast Center Göttingen, Göttingen, Germany.
- Diagnostikum Graz, Graz, Austria.
- Universitätsklinikum Erlangen, Department of Radiology, Erlangen, Germany.
- Medical University of Gdańsk, Second Radiology Department, Gdańsk, Poland.
- National Cancer Center, Department of Radiology, Chuo City, Japan.
- Shinshu University School of Medicine, Department of Radiology, Matsumoto, Japan.
- University Hospital Basel, Basel, Switzerland.
- Dr. D. Y. Patil Vidyapeeth, Dr. D. Y. Patil Medical College, Hospital and Research Center, Pune, Maharashtra, India.
- Sheba Medical Center, Ramat Gan, Israel.
- Diagnostic Mammography Center, Athens, Greece.
- Constructor University, Bremen, Germany.
Abstract
Breast magnetic resonance imaging (MRI) is widely used for diagnostic imaging in breast cancer, creating a need for automated image analysis to assist early detection and diagnosis. We approach the task of lesion suspiciousness classification in breast MRI. Using a multi-centric data collection of 3792 patients, we compare a two-stage classification model based on nnU-Net and a three-dimensional (3D) Swin Transformer with the one-stage classifiers nnDetection and RadYOLO, our 3D extension of YOLO11. In a fivefold leave-sites-out cross-validation scheme, the mean (standard deviation) across folds of the lesion classification area under the receiver operating curve (ROC-AUC) is 0.754 (0.086) for the two-stage model, 0.693 (0.108) for nnDetection, 0.737 (0.074) for RadYOLO, and 0.755 (0.064) for RadYOLO with a separate classification output. Although classification performance does not differ conclusively across methods, RadYOLO achieves 3 to 10 times shorter inference times.