Pose-GuideNet: Automatic Scanning Guidance for Fetal Head Ultrasound from Pose Estimation

Qianhui Men, Xiaoqing Guo, Aris T. Papageorghiou, J. Alison Noble

Introduction

Acquiring high-quality fetal ultrasound images is a skill-intensive task. Sonographers must mentally reconstruct 3D anatomy from 2D cross-sections to locate specific Standard Planes (SPs), such as the trans-ventricular (TVP) and trans-cerebellar (TCP) planes, which are essential for fetal growth assessment. While AI tools exist to detect these planes, they rarely provide guidance on how to reach them. Current navigation solutions often rely on expensive external motion trackers or 3D ultrasound hardware, which are not standard in most clinical settings.

Pose-GuideNet addresses this by providing 3D navigation guidance using only a standard 2D ultrasound probe. It functions as a "GPS" for fetal ultrasound by registering 2D freehand scans to a 3D anatomical atlas without requiring additional sensors.

Methods: A Two-Stage Registration Framework

The core of Pose-GuideNet is a 2D/3D registration approach that aligns freehand 2D ultrasound images to a volumetric 3D fetal head atlas. The architecture utilizes a ResNeXt50-based pose encoder (PoseE) to predict orientation (quaternions) and translation.

The Pose-GuideNet architecture features a two-stage learning process: pre-training on synthetic atlas slices followed by fine-tuning on real-world 2D ultrasound scans using specialized alignment strategies.

The learning process is divided into two stages:

  1. Pre-training: The model is trained on synthetic 2D slices sampled directly from a 3D atlas to establish a baseline mapping between anatomy and 3D space.
  2. Fine-tuning: To adapt to real-world scans, the model employs two novel alignment strategies. Geometric-Guided In-plane Alignment uses supervised learning to align standard planes with known viewpoints in the atlas. Semantic-Aware Out-of-plane Alignment uses an unsupervised contrastive loss to ensure that "searching" frames (those between standard planes) are mapped to anatomically correct locations in the 3D atlas.

To bridge the modality gap between noisy ultrasound and clean atlas slices, the framework uses MedSAM for foreground extraction, focusing the model on the fetal head anatomy and filtering out maternal tissue or shadows.

Key Results

Pose-GuideNet was evaluated on the PULSE dataset, comprising 192 clinical scanning sequences. Ground truth probe orientation was recorded using IMU sensors for validation purposes only.

Quantitative comparison of Pose-GuideNet against baseline methods and sensored systems across TVP and TCP acquisitions.

The results demonstrate that Pose-GuideNet significantly outperforms previous registration methods. Notably, in terms of motion prediction (measured by KL Divergence), Pose-GuideNet (17.33 for TVP) approaches the performance of "sensored" systems (14.79) that utilize hardware trackers. Qualitative results further confirm that retrieved atlas planes closely resemble input ultrasound frames in both anatomy and orientation, even for difficult views like the cerebellum.

Discussion and Future Work

The primary innovation of Pose-GuideNet is its ability to infer 3D spatial context purely from 2D image sequences by leveraging a 3D atlas prior. This effectively turns a standard ultrasound probe into a "smart" device capable of understanding its position within the 3D fetal volume.

While the current model focuses heavily on orientation, future work will aim to improve translation prediction and translate mathematical transformations into "human-readable" instructions (e.g., "Tilt the probe 10 degrees left"). Real-time deployment of these instructions could significantly assist trainee sonographers in clinical practice.

Summary updated

BibTeX
@inproceedings{men2024pose,
  author = {Qianhui Men and Xiaoqing Guo and Aris T. Papageorghiou and J. Alison Noble},
  title = {{Pose-GuideNet: Automatic Scanning Guidance for Fetal Head Ultrasound from Pose Estimation}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2024},
  doi = {10.1007/978-3-031-72083-3\_65},
}
Figure