Abstract
Each medical segmentation task should be considered with a specific AI algorithm based on its scenario so that the most accurate prediction model can be obtained. The most popular algorithms in medical segmentation, 3D U-Net and its variants, can directly implement the task of lung trachea segmentation, but its failure to consider the special tree-like structure of the trachea suggests that there is much room for improvement in its segmentation accuracy. Therefore, a research gap exists because a great amount of state-of-the-art DL algorithms are vanilla 3D U-Net structures, which do not introduce the various performance-enhancing modules that come with special natural image modality in lung airway segmentation. In this paper, we proposed two different network structures Branch-Level U-Net (B-UNet) and Branch-Level CE-UNet (B-CE-UNet) which are based on U-Net structure and compared the prediction results with the same dataset. Specially, both of the two networks add branch loss and central line loss to learn the feature of fine branch endings of the airways. Uncertainty estimation algorithms are also included to attain confident predictions and thereby, increase the overall trustworthiness of our whole model. In addition, predictions of the lung trachea based on the maximum connectivity rate were calculated and extracted during post-processing for segmentation refinement and pruning.
Introduction
Accurate mapping of the human airway tree from 3D CT scans is a critical task in respiratory medicine. Precise segmentation is essential for diagnosing and monitoring conditions such as Chronic Obstructive Pulmonary Disease (COPD) and Idiopathic Pulmonary Fibrosis (IPF). In IPF, the fibrotic patterns often resemble small airways on High-Resolution Computed Tomography (HRCT), making accurate differentiation vital for prognosis.
While 3D U-Net architectures are standard in medical imaging, they often struggle with the intricate, tree-like geometry of pulmonary airways. These models typically treat all voxels with equal weight, frequently failing to capture thin, peripheral branches. Furthermore, the lack of uncertainty quantification in standard models limits their clinical utility, as clinicians cannot assess the reliability of a specific segmentation.
Methodology
To address these challenges, this study introduces two specialized variants of the 3D U-Net: Branch-Level U-Net (B-UNet) and Branch-Level CE-UNet (B-CE-UNet).
B-UNet: Structural Connectivity
B-UNet focuses on preserving the airway's structural integrity through two novel loss functions:
- Branch Loss: This function penalizes the model based on individual branch segments, ensuring that the model prioritizes the detection of the entire tree structure rather than just maximizing volume overlap.
- Central Line Loss: By extracting the "skeleton" of the airway via image erosion, this loss forces the network to maintain the continuity of the airway paths.
B-CE-UNet: Context Extraction
The B-CE-UNet architecture incorporates a Context Extraction (CE) module between the encoder and decoder to capture multi-scale features.
As shown in the architecture overview, the CE module consists of:
- Dense Atrous Convolution (DAC): Uses varying dilation rates to capture features at multiple scales without sacrificing spatial resolution.
- Residual Multi-kernel Pooling (RMP): Employs four different pooling sizes to aggregate global contextual information, helping the model distinguish fine branches from surrounding tissue.
Uncertainty Estimation
To enhance model trustworthiness, Monte Carlo (MC) Dropout was implemented. By performing multiple inferences with random dropout masks, the system generates a variance map alongside the mean prediction, highlighting regions where the model is uncertain.
Experimental Setup
The models were trained and evaluated on a combined dataset of 140 cases from EXACT09 and LIDC-IDRI. The data was partitioned into 72 training, 18 validation, and 50 test cases. Due to memory constraints, 3D patches of size [128, 96, 144] were used.
Results and Discussion
The proposed models were benchmarked against several state-of-the-art (SOTA) architectures, including Attention U-Net and WingsNet. Traditional metrics like the Dice Similarity Coefficient (DSC) were supplemented with structural metrics: Tree Detected (TD) and Branch Detected (BD) ratios.
Key findings include:
- Superior Branch Detection: B-UNet achieved a DSC of 0.897 and a TD ratio of 0.901. B-CE-UNet excelled in the BD ratio (0.884), demonstrating the effectiveness of the context extraction module in identifying fine-grained endings.
- Baseline Comparison: Standard models like V-Net and VoxResNet performed poorly on structural metrics, with TD scores as low as 0.33, compared to ~0.90 for the proposed models.
- Expert Augmentation: Visual analysis revealed that the models often identified subtle branches missed by human experts in the ground truth, suggesting potential for use as a clinical decision support tool.
Conclusion
The introduction of branch-specific loss functions and context extraction modules significantly improves the segmentation of complex airway structures. By incorporating uncertainty estimation, these models provide a more reliable foundation for clinical diagnosis. Future work will focus on reducing disconnected false positives through integrated architectural refinements and developing user interfaces for radiologist verification of low-confidence segments.
Summary updated