Speaker
Description
Recording neuromagnetic fields from the spinal cord with Optically Pumped Magnetometers (OPMs) has the potential to offer new insights into spinal cord function. To realise the full potential of OPM based magnetospinography (OPM-MSG), accurate forward models incorporating subject-specific anatomy are required [1]. These models depend on torso and internal organ geometry (heart, lungs, etc.) but acquiring MR images for each participant is resource-intensive. This hinders large-scale adoption of OPM-MSG.
To investigate whether a template anatomy could be used instead, we propose a deep learning approach to generate individual organ structure from a person’s torso shape. From the TotalSegmentor CT dataset [2], [3], we obtained segmented meshes of the torso and multiple internal tissue types, including the heart, lungs and bones, from 500 participants. We split this randomly into a training set of 440 participants and a test set of 60 participants using 5-fold cross-validation. To enable convolutional network processing, input torsos and organs were converted into fixed-resolution volumetric grids (e.g., 192x192x192 voxels). A conditional 3D U-Net [4], [5] was trained on these pairs, taking the torso voxel grid as input and learning a complex spatial mapping to predict the 3D shape, location, and inter-relation of the internal organs.
To evaluate this template for OPM-MSG, we plan to test two different anatomical templates in simulation. Using the held-out test data from the TotalSegmentor Dataset, we will simulate OPM-MSG data generated from the participants’ own anatomy [1] and reconstruct the source of the simulated data under three conditions: using the participants’ own anatomy, using an existing template warped to the participants’ torso surface and finally using a template generated from the participants’ outer torso surface through a deep learning approach. By considering the distance between the simulated and reconstructed MSG sources in these three conditions, we can begin to answer whether a template could reasonably be used for the spinal cord anatomy in an empirical OPM-MSG recording.
REFERENCES:
[1] G. C. O’Neill, M. E. Spedden, M. Schmidt, S. Mellor, M. Stenroos, and G. R. Barnes, “Volume conductor models for magnetospinography,” Apr. 07, 2025, bioRxiv. doi: 10.1101/2024.11.04.621905.
[2] J. Wasserthal et al., “TotalSegmentator: robust segmentation of 104 anatomical structures in CT images,” Radiol. Artif. Intell., vol. 5, no. 5, p. e230024, Sep. 2023, doi: 10.1148/ryai.230024.
[3] F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,” Nat. Methods, vol. 18, no. 2, pp. 203–211, Feb. 2021, doi: 10.1038/s41592-020-01008-z.
[4] Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ronneberger, “3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation,” Jun. 21, 2016, arXiv: arXiv:1606.06650. doi: 10.48550/arXiv.1606.06650.
[5] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” May 18, 2015, arXiv: arXiv:1505.04597. doi: 10.48550/arXiv.1505.04597.