Model Card for HSF

This model card introduces the Hippocampal Segmentation Factory (HSF), a deep learning model designed for the segmentation of hippocampal subfields in MRI data. HSF leverages an end-to-end deep learning pipeline to provide fast, accurate, and robust segmentation, facilitating the study of volumetric trajectories across the lifespan.

Model Details

Model Description

HSF was developed to address the challenges of segmenting small hippocampal subfields in conventional MRI sequences. It outperforms currently used tools in terms of the Dice Coefficient, Hausdorff Distance, and Volumetric Similarity, making it a valuable tool for neuroimaging research.

Developed by: Clement Poiret, Antoine Bouyeure, Sandesh Patil, Antoine Grigis, Edouard Duchesnay, Matthieu Faillot, Michel Bottlaender, Frederic Lemaitre, Marion Noulhiane.
Funded by: Fondation de France, Grant/Award Number: 00070721.
Model type: 3D Residual U-Net with Self-Attention with modifications for hippocampal subfields segmentation.
License: MIT.

Model Sources

Repository: HSF GitHub Repository.
Paper: Poiret, C., Bouyeure, A., Patil, S., et al. (2023). A fast and robust hippocampal subfields segmentation: HSF revealing lifespan volumetric dynamics. Frontiers in Neuroinformatics, 17. DOI: 10.3389/fninf.2023.1130845.
Documentation: https://hsf.rtfd.io/

Uses

Direct Use

HSF is intended for researchers and clinicians who require accurate segmentation of hippocampal subfields in MRI data. It can be directly used to study volumetric trajectories across the lifespan, aiding in the understanding of episodic memory development and age-related memory impairments.

Downstream Use

HSF's robust segmentation capabilities make it suitable for integration into larger neuroimaging studies, potentially contributing to the development of diagnostic tools for neurological conditions.

Out-of-Scope Use

HSF is not designed for real-time segmentation in clinical settings or for use with non-MRI imaging modalities.

Bias, Risks, and Limitations

HSF was trained and validated on a heterogeneous database comprising all public datasets with manually segmented hippocampal subfields. While it shows superior performance compared to existing tools, its accuracy may vary across different populations or MRI protocols not represented in the training data.

Recommendations

Researchers should validate HSF's performance on their specific datasets, especially when dealing with underrepresented groups or non-standard MRI protocols.

How to Get Started with the Model

Refer to the HSF Documentation for installation instructions, usage examples, and documentation on how to integrate HSF into your neuroimaging pipeline.

Training Details

Training Data

HSF was trained on ~13 datasets of manually segmented hippocampi by individual expert raters, totaling more than 800 subjects.

Training Procedure

HSF employs a bagging technique with five "weak-learner" models to enhance prediction accuracy. It utilizes an AdamP optimizer w/ Adaptive Gradient Clipping and Demon, a one-cycle learning rate scheduler for 64 epochs.

Evaluation

Testing Data, Factors & Metrics

HSF was validated against manual segmentations and compared with ASHS, HIPS, and HippUnfold using the Dice Coefficient, Hausdorff Distance, and Volumetric Similarity.

Results

HSF demonstrated superior performance in terms of segmentation accuracy, with significant improvements over existing tools.

Environmental impact

Experiments were conducted using Scaleway, which has a carbon efficiency of 0.1 kgCO$_2$eq/kWh. A cumulative of 25 hours of computation was performed on hardware of type H100 SXM4 80 GB (TDP of 400W).

Total emissions are estimated to be 1 kgCO$_2$eq of which 100 percents were directly offset by the cloud provider.

Hardware Type: Nvidia H100
Hours used: 25 hours for all models
Cloud Provider: Scaleway
Compute Region: Paris
Carbon Emitted: 1 kgCO2eq

Citation

BibTeX:

@ARTICLE{10.3389/fninf.2023.1130845,
AUTHOR={Poiret, Clement and Bouyeure, Antoine and Patil, Sandesh and Grigis, Antoine and Duchesnay, Edouard and Faillot, Matthieu and Bottlaender, Michel and Lemaitre, Frederic and Noulhiane, Marion},
TITLE={A fast and robust hippocampal subfields segmentation: HSF revealing lifespan volumetric dynamics},
JOURNAL={Frontiers in Neuroinformatics},
VOLUME={17},
YEAR={2023},
URL={https://www.frontiersin.org/articles/10.3389/fninf.2023.1130845},
DOI={10.3389/fninf.2023.1130845},
ISSN={1662-5196},
}