Related Experiment Video
Updated: Jun 21, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Multicenter privacy-preserving model training for deep learning brain metastases autosegmentation.
Yixing Huang1, Zahra Khodabakhshi2, Ahmed Gomaa1
1Department of Radiation Oncology, Universitätsklinikum Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany; Comprehensive Cancer Center Erlangen-EMN (CCC ER-EMN), Erlangen, Germany; Bavarian Cancer Research Center (BZKF), Erlangen, Germany.
Multicenter data heterogeneity challenges deep learning for brain metastases (BM) autosegmentation. Learning without forgetting (LWF) improves model generalizability in privacy-preserving, peer-to-peer training without sharing raw data.
Area of Science:
- Medical imaging analysis
- Artificial intelligence in oncology
- Radiomics and computational pathology
Background:
- Deep learning models for brain metastases (BM) autosegmentation face performance challenges due to data heterogeneity across institutions.
- Variations in imaging protocols, patient populations, and BM characteristics contribute to poor model generalizability.
- Privacy concerns often limit direct data sharing for collaborative model training.
Purpose of the Study:
- To investigate the impact of multicenter data heterogeneity on deep learning-based BM autosegmentation.
- To evaluate the efficacy of an incremental transfer learning technique, learning without forgetting (LWF), for enhancing model generalizability.
- To assess the feasibility of privacy-preserving bilateral collaboration using LWF for model training.
Main Methods:
- Utilized six BM datasets from diverse institutions (UKER, USZ, Stanford, UCSF, NYU, BraTS Challenge 2023).
- Established baseline performance of the DeepMedic network via single-center and mixed multicenter training.
- Evaluated privacy-preserving transfer learning (TL) with and without LWF for peer-to-peer model sharing.
Main Results:
- Single-center training yielded variable F1 scores (0.625-0.876). Mixed multicenter training improved performance at some centers (Stanford, NYU).
- When a model pretrained at UKER was transferred to USZ, LWF achieved a significantly higher F1 score (0.839) compared to naive TL (0.570) and single-center training (0.688).
- LWF demonstrated balanced improvements in sensitivity, precision, and contouring accuracy, whereas naive TL compromised precision.
Conclusions:
- Multicenter data heterogeneity poses significant challenges to the generalizability of deep learning models for BM autosegmentation.
- Learning without forgetting (LWF) is a promising privacy-preserving approach for peer-to-peer model training, enhancing generalizability without raw data sharing.
- LWF effectively addresses data heterogeneity issues, leading to improved and more robust BM autosegmentation performance across different centers.
