Scalp surface estimation and head registration using sparse sampling and 3D statistical models

Oded Schlesinger, Raj Kundu, Dmitry Isaev, Jessica Y. Choi, Stefan M. Goetz, Dennis A. Turner, Guillermo Sapiro, Angel V. Peterchev, J. Matías Di MartinoComputers in Biology and Medicine 2024 · DOI

TL;DR

This paper introduces a novel method for accurately estimating scalp surface shape and registering a subject's head using only sparse, manually acquired surface samples. This approach bypasses the need for patient-specific MRI scans and manual landmark annotation, making head registration more accessible and efficient for applications like neuronavigation and brain stimulation.

What Problem Does This Paper Address?

Accurate head registration and scalp surface estimation are crucial for various biomedical applications, including neuronavigation for brain stimulation and recording. Current methods often rely on manually identified fiducial head targets and require patient-specific MRI scans, which can be time-consuming and limit adoption. The challenge lies in developing a practical and accurate technique that can achieve precise registration and surface estimation without these demanding prerequisites.

Methods

The core of this work is a joint optimization framework that simultaneously performs head shape reconstruction and registration. It leverages a 3D Morphable Model (3DMM), a statistical model representing variations in head shapes, as a prior. The method works by:

  • Sparse Surface Sampling: Acquiring a limited number of points on the scalp surface using existing neuronavigation tools (e.g., a pointer).
  • Statistical Model Prior: Using a pre-existing statistical model of head shapes (the UHM) to infer missing surface information and guide the reconstruction.
  • Joint Optimization: An iterative process that refines both the alignment (registration) of the sampled data to the 3DMM and the deformation of the 3DMM to match the sampled data (shape modeling). This is achieved using deep learning libraries like PyTorch and PyTorch3D for efficient gradient computation and 3D manipulation.
  • Loss Function: A combination of Chamfer distances (measuring vertex and normal similarity between the sampled and reconstructed shapes) and regularization terms to prevent overfitting and ensure smooth transformations.

This approach is novel in its ability to perform accurate reconstruction and registration from sparse and partial data without requiring manual landmark annotation or individual MRIs.

Datasets

The study utilizes two main types of data:

  • MRI Data: Publicly available datasets, IXI (577 subjects) and ADNI1 Complete 1Yr 1.5T (639 subjects), were used for virtual experiments. These datasets provided ground-truth head shapes from T1-weighted and T2-weighted MRI scans.
  • Experimental Data: Data from 50 subjects collected in a clinical lab setting using conventional neuronavigation pointer tools were used for experimental validation.

Key Insights and Ideas

  • Sparse Sampling Sufficiency: The research demonstrates that a surprisingly small number of surface samples (around 2% of the total surface area) can be sufficient for accurate head shape reconstruction and registration.
  • Prior Knowledge is Key: Leveraging statistical head models (3DMMs) as a prior is critical for inferring missing information and achieving high accuracy from sparse data.
  • No Manual Annotation Needed: The method's ability to bypass manual landmark annotation significantly simplifies the workflow and reduces potential sources of error.
  • Sampling Strategy Matters (Initially): While different sampling strategies (e.g., side-to-top, side-to-side) show varying performance at shorter trajectory lengths, their differences diminish as trajectory length increases, suggesting that sufficient coverage is more important than a specific path.
  • Robustness to Noise and Imperfections: The method shows robustness to noisy samples and certain irregularities, making it practical for real-world applications.

Key Results

  • Simulation Accuracy: In virtual studies with 1152 subjects, the proposed method achieved an average reconstruction root-mean-square error (RMSE) of 2.95 mm, outperforming a common neuronavigation technique by 2.70 mm.
  • Experimental Accuracy: On data from 50 subjects collected with conventional tools, the method achieved an average RMSE of 2.89 mm. Adding landmark-based registration further improved this to 2.63 mm.
  • Efficiency: The method is efficient, with reconstruction and registration taking at most 38.1 seconds per subject.
  • Comparison to Benchmark: The proposed method consistently outperformed the Brainsight TMS benchmark, even without manual landmark annotation.

Limitations and Future Work

  • Limited Scope of 3DMM: While the UHM is comprehensive, it might not capture extreme variations in head shape.
  • Sampling Strategy Impact: Although differences diminish with longer trajectories, the choice of sampling strategy can still influence efficiency and data collection effort.
  • Future Work: The authors suggest exploring the application of their method to other anatomical structures and investigating its use in real-time scenarios. Further research into optimizing sampling trajectories for specific clinical needs is also implied.

Summary updated

BibTeX
@article{schlesinger2024scalp,
  author = {Oded Schlesinger and Raj Kundu and Dmitry Isaev and Jessica Y. Choi and Stefan M. Goetz and Dennis A. Turner and Guillermo Sapiro and Angel V. Peterchev and J. Matías Di Martino},
  title = {{Scalp surface estimation and head registration using sparse sampling and 3D statistical models}},
  journal = {Computers in Biology and Medicine},
  year = {2024},
  doi = {10.1016/j.compbiomed.2024.108689},
}
Figure