Keypoint detection network for needle localization on intra-procedural MRI in MRI-guided liver interventions.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Surgical Technology

Background:

• Segmentation neural networks show promise for interventional needle localization on MRI.
• Current segmentation methods require extensive annotation and post-processing, potentially reducing accuracy.
• Needle localization on intra-procedural liver MRI is critical for percutaneous interventions.

Purpose of the Study:

• Develop keypoint detection networks for direct needle entry point and tip localization on intra-procedural liver MRI.
• Improve annotation efficiency and robustness compared to segmentation-based approaches.
• Enable precise needle guidance during liver interventions.

Main Methods:

• Enhanced stacked hourglass models with multi-task learning for keypoint and part affinity field prediction.
• Evaluation on 2D and 3D keypoint detection networks using pre-clinical in vivo pig liver MRI data.
• Comparison with UNet, Swin Transformer segmentation networks, and human intra-reader variation.

Main Results:

• Keypoint networks achieved median localization errors of 1.56 mm (SS-CB) and 2.21 mm (MS-CB).
• Computational times were rapid: 10 ms (2D) and 30 ms (3D).
• Significantly higher accuracy than segmentation models (p < 0.001), comparable to human readers.

Conclusions:

• Keypoint detection networks provide rapid, pixel-level needle localization for intra-procedural liver MRI.
• The proposed method offers higher accuracy and more efficient annotation than segmentation-based models.
• This technology enhances MRI-guided percutaneous liver interventions.