WoundNeRF: Multi-View Consistent Wound Segmentation With Neural Fields

Motivation

Chronic wound care remains a major healthcare challenge, requiring regular assessments that are costly and often delayed.
Most existing automated wound segmentation methods operate in 2D, leading to inconsistent results across viewpoints and inaccurate 3D measurements.

2D annotation inconsistencies observed when projected onto the same 3D surface. Overlapping regions are highlighted in magenta.

WoundNeRF addresses this by learning a 3D‑consistent wound segmentation field directly from multi‑view images.

Approach

Building on advances in Neural Radiance Fields (NeRFs), WoundNeRF jointly models wound appearance, geometry, and semantics in a unified 3D space.

• Aggregates automatically generated 2D segmentations into a 3D semantic field
• Enforces multi‑view consistency without requiring dense manual labeling
• Focuses on accurate 3D reconstruction of wound regions rather than hallucination or blending

WoundNeRF architecture and training pipeline.

Results

We evaluate WoundNeRF on a real patient dataset collected with clinical collaborators.

Baselines:

• 2D: SegFormer MiT‑B5 trained on single‑view annotations
• 3D/2D: Multi‑view aggregation following Wound3DAssist

Our model produces smoother boundaries, higher recall, and stronger multi‑view coherence.

Accuracy comparison across 73 patient videos.

Qualitative comparison of wound and tissue segmentations. Tissues classes: wound bed (1), granulation (2), necrotic (3), slough (4), and unknown (5).

Even with limited expert annotations (2–4 views), WoundNeRF maintains consistent predictions across 50 unseen viewpoints, dramatically reducing inter‑view variation and noise.

Prediction consistency: 2D model (left) vs WoundNeRF (right). The bottom row displays the corresponding wound segmentation on the 3D mesh extracted from our method.

Robustness

When training masks are intentionally perturbed, WoundNeRF remains stable, highlighting strong resistance to noisy supervision.

Robustness under boundary perturbations.

Conclusion

We present WoundNeRF, a method for generating multi‑view consistent wound segmentations.
By learning directly in 3D space, it overcomes the topological limitations of 2D segmentation and enables coherent wound reconstruction for clinical use.

Future work will explore confidence‑driven segmentation to reduce misclassified regions and enhance semantic reliability for real‑world healthcare deployment.

If you find this work useful, please cite:

@misc{chierchia2026woundnerf,
      title={Multi-View Consistent Wound Segmentation With Neural Fields}, 
      author={Remi Chierchia and Léo Lebrat and David Ahmedt-Aristizabal and Yulia Arzhaeva and Olivier Salvado and Clinton Fookes and Rodrigo Santa Cruz},
      year={2026},
      eprint={2601.16487},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.16487}
}

Related Works

If interested, check our previous publications:

• Syn3DWound
  Project Page • Paper (arXiv) • Dataset
• SALVE
  Project Page • Paper (IEEE) • Dataset
• Wound3DAssist
  Paper (arXiv)
• Non-Invasive 3D Wound Measurement with RGB-D Imaging
  Paper (arXiv)

Acknowledgment

This work was supported by the MRFF Rapid Applied Research Translation grant (RARUR000158), CSIRO AI 4 Missions Minimising Antimicrobial Resistance Mission, and Australian Government Training Research Program (AGRTP) Scholarship.