Abstract
Airway segmentation, especially bronchioles segmentation, is an important but challenging task because distal bronchus are sparsely distributed and of a fine scale. Existing neural networks usually exploit sparse topology to learn the connectivity of bronchioles and inefficient shallow features to capture such high-frequency information, leading to the breakage or missed detection of individual thin branches. To address these problems, we contribute a new bronchial segmentation method based on Group Deep Dense Supervision (GDDS) that emphasizes fine-scale bronchioles segmentation in a simple-but-effective manner. First, Deep Dense Supervision (DDS) is proposed by constructing local dense topology skillfully and implementing dense topological learning on a specific shallow feature layer. GDDS further empowers the shallow features with better perception ability to detect bronchioles, even the ones that are not easily discernible to the naked eye. Extensive experiments on the BAS benchmark dataset have shown that our method promotes the network to have a high sensitivity in capturing fine-scale branches and outperforms state-of-the-art methods by a large margin (+12.8 % in BD and +8.8 % in TD) while only introducing a small number of extra parameters.
Introduction
Accurate mapping of the pulmonary airway tree is essential for diagnosing diseases like COPD and asthma. While large airways, such as the trachea and main bronchi, are relatively easy to segment, the peripheral bronchioles—often less than 1mm in diameter—present a significant challenge for medical AI due to their sparse distribution and low contrast against lung tissue.
Current Convolutional Neural Networks (CNNs) often fail in these distal regions for two primary reasons. First, they typically rely on "sparse topology" learning, which focuses on individual points but misses the critical connectivity between branches, resulting in fragmented or "broken" airway trees. Second, standard shallow layers in these networks often lack the sensitivity required to capture high-frequency details, making it difficult to distinguish tiny bronchioles from background noise.
To overcome these limitations, we propose Group Deep Dense Supervision (GDDS). This framework forces the network to perceive local 3D structures early in the feature-extraction process, significantly enhancing the detection of fine-scale branches with minimal computational overhead.
Methods
The GDDS framework introduces a two-pronged strategy to improve segmentation sensitivity and connectivity:
Deep Dense Supervision (DDS)
Instead of traditional voxel-wise classification, DDS transforms ground-truth labels into a local dense topology. By taking a local $n \times n \times n$ cube of voxels and flattening it into $n^3$ channels, a specialized "Deep Dense Head" is trained to predict this local 3D structure from shallow feature layers. This approach forces the network to develop a "dense" understanding of local connectivity and shape rather than looking at isolated pixels.
Group Supervision (GS) and Consistency
To ensure that fine-scale details are preserved as the network goes deeper, GS upsamples and concatenates feature maps from multiple layers (both encoder and decoder) to provide a unified supervision signal. Additionally, a Foreground-Emphasized Consistency Constraint is employed. Because the Deep Dense Head is often more sensitive to tiny airways than the main segmentation head, this constraint forces the main head to align its predictions with the Dense Head in foreground regions, acting as a confidence booster for hard-to-detect branches.
The pipeline also utilizes a generation-aware training paradigm, specifically optimizing the model to handle bronchioles (4th generation and beyond) separately from the larger, more easily segmented airways.
Results
The method was evaluated using the Binary Airway Segmentation (BAS) Dataset, which consists of 90 CT scans with expert-corrected centerlines and branch-level labels. Using UNet and WingsNet as backbones, GDDS significantly outperformed state-of-the-art (SOTA) models.
Quantitative and Qualitative Performance
- Branch Detection (BD): Improved by +12.8% over previous SOTA methods.
- Tree-length Detected (TD): Improved by +8.8%.
- Efficiency: The GDDS module is extremely lightweight, adding only 0.02% extra parameters to the base model.
Visual results demonstrate that GDDS successfully maintains the connectivity of thin branches that were previously fragmented or entirely missed by other architectures.
Conclusion
GDDS offers a simple yet effective solution for segmenting fine-scale pulmonary bronchioles. By emphasizing local dense topology and hierarchical feature perception, the method achieves a substantial leap in branch detection sensitivity. Future work will focus on refining airway boundaries and validating the method across diverse clinical pathologies to support early diagnosis of chronic respiratory conditions.
Summary updated