Iterative Next Boundary Detection for Instance Segmentation of Tree Rings in Microscopy Images of Shrub Cross Sections

Alexander Gillert, Giulia Resente, Alba Anadon-Rosell, Martin Wilmking, Uwe Freiherr von LukasCVPR 2023 · DOI · arXiv

Introduction: The Challenge of Shrub Dendrochronology

Dendrochronology, the study of tree rings, is a vital tool for reconstructing past climate conditions. In the Arctic, shrubs are often the only woody plants available for this analysis. However, analyzing shrub cross-sections is a manual, labor-intensive process. Existing computer vision techniques for "instance segmentation" (identifying individual objects) struggle with tree rings due to their concentric nature, extreme precision requirements, and structural irregularities.

One of the most significant hurdles is the presence of "wedging rings"—rings that do not form a complete circle—and disconnected fragments that break the assumptions of generic segmentation algorithms.

Detailed view of wedging rings (labeled 1-13), illustrating how boundaries can merge or disappear, creating significant challenges for standard automated detection.

Methodology: Iterative Next Boundary Detection (INBD)

To overcome these limitations, researchers developed Iterative Next Boundary Detection (INBD), a specialized "growth-mimicking" AI pipeline. Instead of treating rings as independent objects, INBD follows the natural growth direction of the plant:

  • Initial Semantic Segmentation: A U-Net-based model identifies the "pith" (center), general ring boundaries, and the background.
  • Polar Grid Sampling: The image is "unrolled" into a rectangular strip using polar coordinates. This transforms complex circular geometry into a simpler boundary-tracking task.
  • Iterative Detection: Starting from the center, the model detects the first ring boundary and uses it as the starting point to look for the next, repeating the process until it reaches the bark.
  • Wedging Ring Detection (WRD): A recurrent module tracks whether a ring is "opening" or "closing," allowing the model to handle incomplete circles that generic methods often skip or merge.
  • Iterative Training: The model is trained on its own previous predictions, making it robust to errors that might otherwise propagate through the sequence of rings.

Key Results and Insights

INBD was tested on a new, publicly available dataset of 213 high-resolution microscopy images across three species: Dryas octopetala (DO), Empetrum hermaphroditum (EH), and Vaccinium myrtillus (VM).

Comparison between original shrub cross-section microscopy (left) and the resulting instance segmentation masks (right) produced by the INBD pipeline.

The model consistently outperformed both top-down (Mask R-CNN) and bottom-up (GASP, Multicut) methods:

  • Superior Recall: On the EH subset, INBD achieved a Mean Average Recall (mAR) of 0.738, dwarfing Mask R-CNN (0.176).
  • Chronological Ordering: Unlike generic segmentation, INBD has an inherent understanding of time. Because it detects rings sequentially, it automatically assigns a chronological order essential for climate reconstruction.
  • Precision: INBD was the only method capable of reliably separating rings that were very close together or had faint boundaries.

Limitations and Future Work

While INBD is a significant advancement, challenges remain. A major failure in an early ring can still occasionally derail the detection of outer rings. Future research will focus on weakly supervised training to handle the high cost of manual annotation and improving cross-species generalization, as different shrubs possess vastly different cellular structures.

Summary updated

BibTeX
@inproceedings{gillert2023iterative,
  author = {Alexander Gillert and Giulia Resente and Alba Anadon-Rosell and Martin Wilmking and Uwe Freiherr von Lukas},
  title = {{Iterative Next Boundary Detection for Instance Segmentation of Tree Rings in Microscopy Images of Shrub Cross Sections}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2023},
  doi = {10.1109/cvpr52729.2023.01397},
}
Figure