SMILEUHURA_DS6_CamSVD_UNetMSS3D_wDeform

The main model from DS6 (UNetMSS + Deformation-aware learning), first trained on 7T ToF-MRAs of healthy subjects from the SMILE-UHURA dataset (https://doi.org/10.7303/syn47164761) and was then fine-tuned on the Cambridge 7T Cerebral Small Vessel Disease (CamSVD) dataset containing data from subjects with Lacunar strokes with SVD, non-lacunar strokes without SVD, and healthy controls.

This study aimed to develop a semi-automatic pipeline for quantifying the 3D morphology of lenticulostriate arteries (LSAs) in small vessel disease (SVD) patients using 7 Tesla Time-of-Flight magnetic resonance angiography (TOF-MRA). The motivation behind the work stems from the limitations of manual 2D analysis methods, which fail to capture the full complexity of LSA morphology. To address this, the authors fine-tuned the deep learning-based DS6 model for vessel segmentation and compared its performance to the classical Frangi filter-based MSFDF method. The DS6 model outperformed MSFDF with a mean Dice similarity coefficient of 0.814±0.029, showing superior sensitivity in detecting weaker LSA branches. Quantifying the segmented LSAs revealed various metrics, including the number of branches, length, and tortuosity. This deep learning-based pipeline offers a more accurate and efficient approach for assessing the 3D structure of LSAs, facilitating future research into the pathophysiology of SVD.

Model Details

It was introduced in DS6, Deformation-Aware Semi-Supervised Learning: Application to Small Vessel Segmentation with Noisy Training Data by Soumick Chatterjee, Kartik Prabhu, Mahantesh Pattadkal, Gerda Bortsova, Chompunuch Sarasaen, Florian Dubost, Hendrik Mattern, Marleen de Bruijne, Oliver Speck, Andreas Nürnberger. ArXiv preprint

The model architecture is the same as the original paper, but it was first trained on the SMILE-UHURA dataset, and then fine-tuned on the Cambridge 7T Cerebral Small Vessel Disease (CamSVD) dataset. While fine-tuning, additional 3D Spatial Dropout layers were added within the convolutional blocks. The results of this fine-tuned model are presented in A Deep Learning Based Pipeline for the Analysis of the 3D Morphology of Lenticulostriate Arteries in Cerebral Small Vessel Disease from 7 Tesla Time-of-Flight MRA

Model Description

Model type: UNet Multi-scale Supervision (UNet-MSS) 3D
Task: Vessel segmentation in 7T MRA-ToF volumes
Initial training dataset: 7T ToF-MRAs from the vessel segmentation challenge: SMILE-UHURA (https://doi.org/10.7303/syn47164761)
Fine-tuning dataset: 7T ToF-MRAs from the Cambridge 7T Cerebral Small Vessel Disease (CamSVD) study, consisting of healthy controls, and subjects with lacunar strokes with SVD and non-lacunar strokes without SVD.
Training type: Trained with deformation-aware learning

Model Sources

DS6 (vessel segmentation):

Repository: https://github.com/soumickmj/DS6
Paper: https://doi.org/10.3390/jimaging8100259
Preprint: https://arxiv.org/abs/2006.10802

This extended study (post-vessel segmentation):

Repository: ComingSoon!
Preprint: ComingSoon!

Citation

If you use this approach in your research or use codes from this repository or these weights, please cite all the following in your publications:

BibTeX:

DS6:

@article{chatterjee2022ds6,
  title={Ds6, deformation-aware semi-supervised learning: Application to small vessel segmentation with noisy training data},
  author={Chatterjee, Soumick and Prabhu, Kartik and Pattadkal, Mahantesh and Bortsova, Gerda and Sarasaen, Chompunuch and Dubost, Florian and Mattern, Hendrik and de Bruijne, Marleen and Speck, Oliver and N{\"u}rnberger, Andreas},
  journal={Journal of Imaging},
  volume={8},
  number={10},
  pages={259},
  year={2022},
  publisher={MDPI}
}

APA:

Chatterjee, S., Prabhu, K., Pattadkal, M., Bortsova, G., Sarasaen, C., Dubost, F., ... & Nürnberger, A. (2022). Ds6, deformation-aware semi-supervised learning: Application to small vessel segmentation with noisy training data. Journal of Imaging, 8(10), 259.