Dealing with Segmentation Errors in Needle Reconstruction for MRI-Guided Brachytherapy

Vangelis Kostoulas, Arthur Guijt, Ellen M. Kerkhof, Bradley R. Pieters, Peter A. N. Bosman, Tanja AlderliestenarXiv 2025 · DOI · arXiv

Abstract

Automated needle reconstruction in MRI-guided brachytherapy is often compromised by segmentation artifacts such as phantom needles and disconnected segments. This paper presents robust post-processing refinements, termed MJung+ and Leon+, which utilize RANSAC-based merging and iterative pruning to rectify these errors. Evaluated on clinical prostate cancer datasets, the proposed methods achieve sub-millimeter shaft accuracy and eliminate false positives, bridging the gap between raw AI segmentation and clinical utility.

Introduction

In interstitial brachytherapy, clinicians treat cancer by placing radioactive sources directly into tumor tissue via specialized needles. Accurate radiation dose planning requires a precise 3D reconstruction of the needle geometry. Manually identifying these needles on MRI scans is a tedious and time-consuming process, creating a significant bottleneck in the clinical workflow.

While deep learning models like nnU-Net have shown promise in segmenting needles, they frequently produce "messy" outputs characterized by false positives (phantom fragments), under-segmentation, and disconnections where a single needle appears as several broken segments. Existing post-processing algorithms often fail under these conditions because they assume clean data or specific anatomical starting points. This research introduces robust adaptations to existing methods to handle these common AI errors.

Overview of the needle reconstruction pipeline, from raw MRI slices to final post-processed needle models.

Methods

The proposed solution employs a two-stage pipeline: automated segmentation followed by robust post-processing.

Segmentation

The authors utilized nnU-Net, comparing 2D and 3D configurations. They found that 3D models significantly reduced the initial number of disconnected segments and false positives compared to 2D counterparts, even with limited training data. The 3D geometry helps the model maintain spatial consistency across slices.

Robust Post-Processing (MJung+ and Leon+)

The core contribution lies in the "plus" versions of two existing methods: Jung (2D clustering + Expectation Maximization) and Leon (3D clustering via HDBSCAN). Key enhancements include:

  • Modified Jung (MJung): Unlike the original version, MJung removes the assumption that needles must start on a specific axial slice. It uses 3D convolutions to identify endpoints and clusters from both ends of the needle, increasing flexibility for varied needle placements.
  • Needle Merging via RANSAC: To address disconnections, the RANSAC algorithm identifies whether disconnected segments belong to the same polynomial curve. Segments within a 5-voxel threshold are merged into a single entity.
  • Iterative Removal: To eliminate false positives, the algorithm uses a "survival of the fittest" approach. If more needles are detected than were used in the procedure, it iteratively prunes the segment that contributes least to the overall model fit until the count is correct.
The post-processing stage specifically targets the refinement of fragmented or noisy segmentation outputs into coherent needle trajectories.

Results

The study was validated using two clinical datasets from the Amsterdam University Medical Center:

  1. Dataset 1: 20 MRI scans (turbo spin echo) containing 293 Ella-CS 6F needles, used for training and cross-validation.
  2. Dataset 2: 14 MRI scans containing 261 Nucletron ProGuide needles, serving as an independent test set.

The MJung+ and Leon+ methods demonstrated significant improvements over baseline techniques. When paired with 3D segmentation on the independent test set, the methods achieved:

  • High Accuracy: A median shaft error of 0.75 mm and a needle-tip localization error of 1.07 mm.
  • Perfect Reliability: Both "plus" methods achieved zero false positives and zero false negatives, successfully reconstructing all 261 needles in the test set. In contrast, the original Jung method missed several needles due to its inability to handle varied starting positions or fragmented segments.

Discussion and Conclusion

The results indicate that 3D geometric information is crucial for reducing initial segmentation noise. Furthermore, the inclusion of RANSAC-based merging and loss-based pruning allows automated systems to overcome the inherent imperfections of AI segmentation.

While the geometric accuracy is sub-millimeter, future work should focus on formal clinical validation to assess the impact on radiation dose planning. Additionally, while the current "Iterative Removal" step requires knowing the needle count (standard clinical information), future iterations could aim for fully autonomous count determination. This work represents a significant step toward a fully automated, reliable clinical workflow for MRI-guided brachytherapy.

Summary updated

BibTeX
@misc{kostoulas2025dealing,
  author = {Vangelis Kostoulas and Arthur Guijt and Ellen M. Kerkhof and Bradley R. Pieters and Peter A. N. Bosman and Tanja Alderliesten},
  title = {{Dealing with Segmentation Errors in Needle Reconstruction for MRI-Guided Brachytherapy}},
  year = {2025},
  eprint = {2507.18895},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.18895},
}
Figure