Deep-Learning for Tidemark Segmentation in Human Osteochondral Tissues Imaged with Micro-computed Tomography

Aleksei Tiulpin, Mikko Finnilä, Petri Lehenkari, Heikki J. Nieminen, Simo Saarakkala

Introduction

Osteoarthritis (OA) is a progressive and debilitating disease characterized by the breakdown of articular cartilage. To study the structural changes associated with OA in three dimensions, researchers frequently utilize micro-computed tomography (μCT). However, soft tissues like cartilage are nearly radiolucent, necessitating the use of contrast agents such as phosphotungstic acid (PTA) to make them visible under X-ray.

While PTA successfully enhances cartilage, it introduces a significant technical challenge: the agent accumulates heavily at the "tidemark"—the critical boundary between calcified and non-calcified cartilage. This accumulation creates a "blind spot" where the boundary effectively disappears, making manual or threshold-based segmentation of the calcified cartilage interface (CCI) nearly impossible. Furthermore, non-enhancing "voids" within the tissue further complicate automated analysis.

Figure 1: Examples of μCT slices showing (a) loss of contrast at the tidemark and (b) non-enhancing "voids" that complicate segmentation.

Methodology

To overcome these imaging artifacts, this study proposes a deep-learning approach using a U-Net architecture. The primary hurdle for such a model is the lack of reliable ground truth data, as the tidemark is obscured in standard PTA scans.

Cross-Modality Ground Truth Generation

The researchers developed a novel training strategy using two different contrast agents. Samples were first scanned using CA4+, a contrast agent that clearly delineates the tidemark. After generating segmentation masks from these scans, the samples were washed and re-scanned using PTA. By co-registering the 3D volumes from both scans, the team produced high-accuracy ground truth masks for the PTA-stained images, effectively teaching the model to "see" the obscured boundary.

Figure 3: The data processing pipeline, showing how CA4+ and PTA scans are co-registered to train the U-Net.

Model Optimization

The U-Net was trained on 2D slices extracted from the ZX and ZY planes of the 3D volumes to maximize the available data. The researchers compared several loss functions, ultimately finding that a combined Binary Cross-Entropy (BCE) and logarithmic Jaccard loss was superior. This combination specifically improved the model's ability to handle the complex, thin geometry of the tidemark interface.

Results and Discussion

The study utilized 35 human articular cartilage samples from patients undergoing total knee replacement. The deep-learning model demonstrated high robustness, particularly in its ability to identify the tidemark despite the presence of PTA-induced voids and contrast loss.

Performance was evaluated using Intersection over Union (IoU) within specific proximity zones of the tidemark:

  • Tight Zone (15 μm): The model achieved an IoU of 0.59.
  • Extended Zone (75 μm): The IoU improved to 0.86.

The stability of metrics like Dice and Volumetric Similarity across widening evaluation zones suggests that the model successfully captured the overall topology of the osteochondral interface, even where local contrast was poorest.

Conclusion

This automated deep-learning framework provides a scalable solution for segmenting the tidemark in high-resolution μCT scans, bypassing the need for labor-intensive manual annotation or expensive double-imaging protocols. By making their code and dataset publicly available, the authors aim to accelerate research into early-stage osteoarthritis and the structural mechanics of the calcified cartilage interface. Future work may involve transitioning to a 3D U-Net architecture to further leverage volumetric context and exploring surface-related metrics like Hausdorff distance for finer anatomical precision.

Summary updated

Abstract

Three-dimensional (3D) semi-quantitative grading of pathological features in articular cartilage (AC) offers significant improvements in basic research of osteoarthritis (OA). We have earlier developed the 3D protocol for imaging of AC and its structures which includes staining of the sample with a contrast agent (phosphotungstic acid, PTA) and a consequent scanning with micro-computed tomography. Such a protocol was designed to provide X-ray attenuation contrast to visualize AC structure. However, at the same time, this protocol has one major disadvantage: the loss of contrast at the tidemark (calcified cartilage interface, CCI). An accurate segmentation of CCI can be very important for understanding the etiology of OA and ex-vivo evaluation of tidemark condition at early OA stages. In this paper, we present the first application of Deep Learning to PTA-stained osteochondral samples that allows to perform tidemark segmentation in a fully-automatic manner. Our method is based on U-Net trained using a combination of binary cross-entropy and soft-Jaccard loss. On cross-validation, this approach yielded intersection over the union of 0.59, 0.70, 0.79, 0.83 and 0.86 within 15 \(\upmu \)m, 30 \(\upmu \)m, 45 \(\upmu \)m, 60 \(\upmu \)m. and 75 \(\upmu \)m padded zones around the tidemark, respectively. Our codes and the dataset that consisted of 35 PTA-stained human AC samples are made publicly available together with the segmentation masks to facilitate the development of biomedical image segmentation methods.
BibTeX
@inproceedings{tiulpin2020deep,
  author = {Aleksei Tiulpin and Mikko Finnilä and Petri Lehenkari and Heikki J. Nieminen and Simo Saarakkala},
  title = {{Deep-Learning for Tidemark Segmentation in Human Osteochondral Tissues Imaged with Micro-computed Tomography}},
  booktitle = {Acivs},
  year = {2020},
  doi = {10.1007/978-3-030-40605-9\_12},
}
Figure