Break: Bronchi Reconstruction by Geodesic Transformation and Skeleton Embedding

Weihao Yu, Hao Zheng, Minghui Zhang, Hanxiao Zhang, Jiayuan Sun, Jie Yang

TL;DR

Convolutional neural networks often struggle to map the human airway in CT scans, frequently predicting broken or disconnected branches due to the tiny size of peripheral bronchi. This paper introduces a multi-stage framework that uses a novel "breakage-sensitive" loss function and geodesic distance transforms to maintain the topological continuity of the airway tree. The result is a highly accurate segmentation model that captures significantly more small branches without requiring tedious manual repair.

What Problem Does This Paper Address?

Mapping the bronchial tree from CT scans is a critical first step for virtual bronchoscopy, surgical navigation, and diagnosing lung diseases. While CNNs are the go-to tool for this 3D segmentation task, their outputs often suffer from "breakages"—discontinuous branches that render the 3D model broken.

These breaks happen for three main reasons:

  1. Intra-class imbalance: Peripheral bronchi are microscopic compared to the main airways. Because standard loss functions are volume-based, a network isn't heavily penalized for missing these tiny distal branches.
  2. Vessel interference: Nearby blood vessels can have similar or higher densities (Hounsfield units) than airway walls, confusing the model's boundary detection.
  3. Pathology: Lung diseases and lesions can deform the airway walls, creating irregular shapes and thicknesses that the network struggles to recognize.

Methods

The authors propose a two-stage, multi-branch architecture to tackle these challenges:

  • Geodesic Distance Transform (GDT): While airway walls change appearance due to disease or nearby vessels, the lumen (the hollow inside of the airway) is consistently dark. The authors use GDT to calculate the shortest path from centerline voxels to other voxels based on image intensity. This effectively traces the dark tubes of the airway, ignoring deformed walls.
  • Breakage-Sensitive (BS) Loss: Standard loss functions (like Dice or Cross-Entropy) care about volume. Missing a tiny branch barely affects the total loss. BS loss fixes this by calculating loss based only on the airway's centerline (its topological skeleton). This applies a massive penalty if a branch breaks, regardless of how thin the branch is.
  • Skeleton Embedding Fusion: The framework trains two branches in parallel—one fine-tuned for high precision, and one using GDT for high continuity. A fusion algorithm then takes the unbroken centerlines from the GDT branch and embeds them into the high-precision output, patching any gaps in the final prediction.

Datasets

  • Binary Airway Segmentation Dataset (BASD): A public dataset of 90 CT images (70 from the LIDC dataset and 20 from the EXACT'09 training set).
  • In-house Dataset: 50 thin-slice CT images collected by the authors and annotated by five experienced radiologists.

Key Insights and Ideas

  • Topology over Volume: By shifting the loss function's focus from the total volume of the airway to its 1D centerline (BS loss), the network is forced to care about the connectivity of the smallest peripheral branches just as much as the main trachea.
  • Intensity-guided routing beats spatial routing: Using Geodesic distance (which factors in pixel intensity) rather than simple Euclidean distance allows the algorithm to follow the natural, dark path of the airway lumen. This makes it highly robust against wall deformations caused by lesions.
  • The "False Positive" Paradox: The model became so good at finding tiny, faint airways that it detected branches the human expert annotators missed. In standard evaluation metrics, these are penalized as "false positives," even though retrospective clinical review confirmed they were actual, true airways.

Key Results

The proposed method achieved state-of-the-art performance, particularly in its ability to detect more of the airway tree without fracturing:

  • On the public BASD dataset, it achieved a 95.33% tree length detected rate and a 94.22% branch detected rate, outperforming baseline methods by roughly 2% to 5%.
  • On the in-house dataset, it reached 92.21% length detected and 90.06% branch detected.
  • Ablation studies (specifically Figure 1 in the paper) showed that the BS loss alone is vastly more sensitive to structural fractures than standard losses (Dice, Cross-Entropy, etc.), reacting sharply the moment a branch breaks during training.

Limitations and Future Work

  • Precision vs. Sensitivity Trade-off: The hyper-focus on finding tiny branches via BS loss leads to a slight drop in mathematical precision. While some of this is due to the "false positive paradox" mentioned above, the model's aggressiveness can occasionally cause slight leakage into background textures that look like fading airway walls.
  • Sensitivity to Noise: The authors noted that calculating geodesic distance using image gradients is highly sensitive to noise and can be disturbed by lesions. They mitigated this by using raw gray values instead, but handling extreme noise remains a challenge.
  • Annotation Difficulty: Because the model is now outperforming human annotators on the smallest branches, future work will likely require better, perhaps AI-assisted, ground-truth generation to accurately evaluate these highly sensitive models.

Summary updated

Abstract

Airway segmentation is critical for virtual bronchoscopy and computer-aided pulmonary disease analysis. In recent years, convolutional neural networks (CNNs) have been widely used to delineate the bronchial tree. However, the segmentation results of the CNN-based methods usually include many discontinuous branches, which need manual repair in clinical use. A major reason for the breakages is that the appearance of the airway wall can be affected by the lung disease as well as the adjacency of the vessels, while the network tends to overfit these special patterns in the training set. To learn robust features for these areas, we design a multi-stage framework that adopts the geodesic distance transform to capture the intensity changes between airway lumen and wall. Another reason for the breakages is the intra-class imbalance. Since the volume of the peripheral bronchi may be much smaller than the large branches in an input patch, the common segmentation loss is not sensitive to the breakages among the distal branches. Therefore, in this paper, a breakage-sensitive regularization term is designed and can be easily combined with other loss functions. Extensive experiments are conducted on publicly available datasets. Compared with state-of-the-art methods, our framework can detect more branches while maintaining competitive segmentation performance.
BibTeX
@inproceedings{yu2022break,
  author = {Weihao Yu and Hao Zheng and Minghui Zhang and Hanxiao Zhang and Jiayuan Sun and Jie Yang},
  title = {{Break: Bronchi Reconstruction by Geodesic Transformation and Skeleton Embedding}},
  booktitle = {ISBI},
  year = {2022},
  url = {https://openreview.net/forum?id=O459gVYfzX},
}

Loading related content...

Figure