Rapid head-pose detection for automated slice prescription of fetal-brain MRI

Malte Hoffmann, Esra Abaci Turk, Borjan A. Gagoski, Leah Morgan, Paul Wighton, M. Dylan Tisdall, Martin Reuter, Elfar Adalsteinsson, Patricia Ellen Grant, Lawrence L. Wald, André J. W. van der Kouwe

Abstract

In fetal-brain MRI, head-pose changes between prescription and acquisition present a challenge to obtaining the standard sagittal, coronal and axial views essential to clinical assessment. As motion limits acquisitions to thick slices that preclude retrospective resampling, technologists repeat ~55-second stack-of-slices scans (HASTE) with incrementally reoriented field of view numerous times, deducing the head pose from previous stacks. To address this inefficient workflow, we propose a robust head-pose detection algorithm using full-uterus scout scans (EPI) which take ~5 seconds to acquire. Our ~2-second procedure automatically locates the fetal brain and eyes, which we derive from maximally stable extremal regions (MSERs). The success rate of the method exceeds 94% in the third trimester, outperforming a trained technologist by up to 20%. The pipeline may be used to automatically orient the anatomical sequence, removing the need to estimate the head pose from 2D views and reducing delays during which motion can occur.

Introduction: The Challenge of "Chasing the Fetus"

Fetal-brain MRI is a vital tool for diagnosing developmental issues, but it faces a significant "moving target" problem. To obtain clear diagnostic images, radiologists require specific anatomical views: sagittal, coronal, and axial. However, fetuses rarely remain stationary during the imaging process.

Currently, MRI technologists employ a manual workflow often described as "chasing the fetus." They acquire a 55-second stack of slices (HASTE), observe where the head has moved, manually re-orient the scanner's field of view, and attempt the scan again. This process is highly inefficient; clinical data indicates that roughly 25% of scans have incorrect orientation, and nearly 38% are severely compromised by motion. This leads to prolonged exam times, patient discomfort, and potential diagnostic delays.

Comparison of misaligned (A) and correctly aligned (C) fetal MRI views, with the low-resolution scout (B) used for automated detection.

Methodology: Leveraging Traditional Computer Vision

The proposed solution is a fully automated pipeline that avoids the need for expensive GPUs or slow deep-learning preprocessing. The approach relies on four key pillars:

  1. Fast Scout Scans: The system utilizes Echo Planar Imaging (EPI), which captures the full uterus in approximately 5 seconds.
  2. MSER Detection: The core of the algorithm uses Maximally Stable Extremal Regions (MSERs). These are image "blobs" that remain stable across different intensity thresholds, making them robust to the varying contrast levels found in fetal MRI.
  3. Anatomical Priors: Instead of a "black box" model, the algorithm filters MSER blobs based on the known size and shape of a fetal brain and eyes at a specific gestational age (GA). By using GA as an input parameter, the algorithm dynamically adjusts its expectations for brain size.
  4. Coordinate System Construction: Once the brain and eyes are located, the algorithm calculates a 3D coordinate system (left-right, posterior-anterior, and inferior-superior) to provide the MRI scanner with precise orientation instructions for the subsequent high-resolution diagnostic scan.

Results: Speed and Accuracy in the Third Trimester

The study evaluated the algorithm across several datasets, including clinical surveys of human technologists, healthy third-trimester fetuses, and fetuses with neurological abnormalities (e.g., ventriculomegaly).

  • High Accuracy: The algorithm successfully localized the brain and eyes in 95.1% of healthy third-trimester scouts.
  • Speed: The entire detection process takes approximately 2 seconds on a standard single-threaded CPU, making it suitable for real-time clinical use without specialized hardware.
  • Clinical Superiority: The automated method outperformed trained technologists by up to 20% in terms of prescription success rate. Human technologists only achieved a "perfect" orientation about 75% of the time in the clinical survey.
  • Robustness: The system maintained a 90% success rate even in cases with significant neurological abnormalities, such as enlarged fluid spaces in the brain.

Discussion and Future Directions

While highly effective in the third trimester, the primary limitation of the current model is its performance in early pregnancy. In the second trimester (14–21 weeks), the success rate dropped to 18.2%. This is attributed to the less developed fetal skull providing lower contrast against amniotic fluid and the smaller brain size being harder to distinguish from maternal tissues.

Future work aims to improve detection in younger fetuses through higher-resolution scout scans or hybrid modeling. Integrating this "auto-pilot" directly into scanner software could allow for continuous, real-time tracking of the fetus throughout an entire MRI session, significantly streamlining fetal neuroimaging.

Summary updated

Abstract

In fetal-brain MRI, head-pose changes between prescription and acquisition present a challenge to obtaining the standard sagittal, coronal and axial views essential to clinical assessment. As motion limits acquisitions to thick slices that preclude retrospective resampling, technologists repeat ~55-second stack-of-slices scans (HASTE) with incrementally reoriented field of view numerous times, deducing the head pose from previous stacks. To address this inefficient workflow, we propose a robust head-pose detection algorithm using full-uterus scout scans (EPI) which take ~5 seconds to acquire. Our ~2-second procedure automatically locates the fetal brain and eyes, which we derive from maximally stable extremal regions (MSERs). The success rate of the method exceeds 94% in the third trimester, outperforming a trained technologist by up to 20%. The pipeline may be used to automatically orient the anatomical sequence, removing the need to estimate the head pose from 2D views and reducing delays during which motion can occur.
BibTeX
@article{hoffmann2021rapid,
  author = {Malte Hoffmann and Esra Abaci Turk and Borjan A. Gagoski and Leah Morgan and Paul Wighton and M. Dylan Tisdall and Martin Reuter and Elfar Adalsteinsson and Patricia Ellen Grant and Lawrence L. Wald and André J. W. van der Kouwe},
  title = {{Rapid head-pose detection for automated slice prescription of fetal-brain MRI}},
  journal = {Int. J. Imaging Syst. Technol.},
  year = {2021},
  doi = {10.1002/ima.22563},
}
Figure