ASCHOPLEX encounters Dafne: a federated continuous learning project for the generalizability of the Choroid Plexus automatic segmentation
Abstract
The Choroid Plexus (ChP) is a highly vascularized brain structure that plays a critical role in several physiological processes. ASCHOPLEX, a deep learning-based segmentation toolbox with an integrated fine-tuning stage, provides accurate ChP delineations on non-contrast-enhanced T1-weighted MRI scans; however, its performance is hindered by inter-dataset variability. This study introduces the first federated incremental learning approach for automated ChP segmentation from 3D T1-weighted brain MRI, by integrating an enhanced version of ASCHOPLEX within the Dafne (Deep Anatomical Federated Network) framework. A comparative evaluation is conducted to assess whether federated incremental learning through Dafne improves model generalizability across heterogeneous imaging conditions, relative to the conventional fine-tuning strategy employed by standalone ASCHOPLEX. The experimental cohort comprises 2,284 subjects, including individuals with Multiple Sclerosis as well as healthy controls, collected from five independent MRI datasets. Results indicate that the fine-tuning strategy provides high performance on homogeneous data (e.g., same MRI sequence, same cohort of subjects), but limited generalizability when the data variability is high (e.g., multiple MRI sequences, multiple and new cohorts of subjects). By contrast, the federated incremental learning variant of ASCHOPLEX constitutes a robust alternative consistently achieving higher generalizability and more stable performance across diverse acquisition settings.
Summary
This paper addresses the challenge of generalizability in automated segmentation of the Choroid Plexus (ChP) from 3D T1-weighted brain MRI using deep learning. The existing ASCHOPLEX toolbox, while accurate, suffers from performance degradation when applied to datasets with different imaging characteristics. To overcome this, the authors introduce a federated incremental learning approach by integrating ASCHOPLEX within the Dafne framework. They compare this new approach against the standard fine-tuning strategy used by standalone ASCHOPLEX. The study utilizes a large cohort of 2,284 subjects from five independent MRI datasets, including healthy controls and individuals with Multiple Sclerosis. The key finding is that federated incremental learning through Dafne consistently achieves higher generalizability and more stable performance across diverse acquisition settings compared to fine-tuning. This matters to the field because it provides a privacy-preserving and robust method for improving the reliability of automated medical image segmentation in multi-center studies where data heterogeneity is a significant concern.
Key Insights
- •The paper demonstrates that fine-tuning ASCHOPLEX on specific datasets leads to high performance on those datasets but poor generalizability to others, highlighting the issue of catastrophic forgetting. For instance, Model FT1 (fine-tuned on MS_Lesion) achieved a Dice score of 0.695 on MS_Lesion but only 0.580 on Verona.
- •Federated incremental learning with Dafne mitigates catastrophic forgetting, maintaining stable performance on the initial dataset (Verona) while improving performance on subsequent datasets. The Dice score on Verona remained consistently between 0.792 and 0.817 across all incremental learning steps.
- •The integration of ASCHOPLEX with Dafne extends Dafne's capabilities to 3D segmentation and introduces support for PyTorch and MONAI, enhancing the platform's versatility.
- •The study incorporates the SwinUNETR architecture into ASCHOPLEX, demonstrating the value of swin transformers for ChPV segmentation. The selected DNN configurations included two SwinUNETR architectures out of the five best models.
- •The 'General' dataset, created to evaluate generalizability, revealed that Model IL4 (the final federated incremental learning model) achieved the highest median Dice value (0.788) with reduced variability, indicating improved cross-dataset generalization compared to Model 0 (0.765).
- •A limitation of the study is that it only considered data from 3T MRI scanners and two manufacturers (Philips and Siemens), potentially limiting the generalizability of the findings to other imaging environments.
- •The study shows that while fine-tuning can achieve higher performance on a single target dataset (e.g., Model FT1 on MS_Lesion), federated incremental learning offers a better trade-off between performance and generalizability across diverse datasets (e.g., Model IL4 on the General dataset).
Practical Implications
- •The federated incremental learning approach can be applied to other medical image segmentation tasks, particularly in scenarios where data is distributed across multiple institutions and data privacy is a concern.
- •Researchers and engineers can leverage the Dafne framework and the integrated ASCHOPLEX toolbox to develop and deploy privacy-preserving and generalizable segmentation models for various anatomical regions.
- •The findings suggest that practitioners should consider federated incremental learning as a robust alternative to fine-tuning when dealing with heterogeneous datasets from multi-center studies.
- •The open-source availability of the ASCHOPLEX-Dafne integration on GitHub facilitates further research and development in federated learning for medical imaging.
- •Future research should focus on expanding the training data to include images from different MRI field strengths (e.g., 1.5T, 7T) and more vendors to further improve the generalizability of the segmentation models.