⚡ electrical engineering

Bronchovascular Tree-Guided Weakly Supervised Learning Method for Pulmonary Segment Segmentation

This paper proposes Anatomy-Hierarchy Supervised Learning (AHSL), a weakly supervised method that leverages bronchovascular tree priors and a two-stage strategy with hierarchical loss functions to achieve accurate pulmonary segment segmentation without requiring laborious pixel-wise annotations.

Original authors: Ruijie Zhao, Zuopeng Tan, Xiao Xue, Longfei Zhao, Bing Li, Zicheng Liao, Ying Ming, Jiaru Wang, Ran Xiao, Sirong Piao, Rui Zhao, Qiqi Xu, Wei Song

Published 2026-04-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ruijie Zhao, Zuopeng Tan, Xiao Xue, Longfei Zhao, Bing Li, Zicheng Liao, Ying Ming, Jiaru Wang, Ran Xiao, Sirong Piao, Rui Zhao, Qiqi Xu, Wei Song

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your lungs aren't just a single, giant balloon, but a complex city made of 18 distinct neighborhoods (called segments). Doctors need to map these neighborhoods perfectly to plan surgeries or treat cancer. If they remove the wrong neighborhood, it could be disastrous.

The problem? In a standard CT scan (a 3D X-ray), these neighborhoods don't have clear fences or walls. They blend into each other like watercolors. It's incredibly hard for a computer to draw the lines between them because there are no visible borders.

This paper introduces a clever new AI method called AHSL (Anatomy-Hierarchy Supervised Learning) to solve this puzzle. Here is how it works, broken down into simple concepts:

1. The Problem: The "Invisible Fence"

Usually, to teach a computer to draw a map, you show it thousands of examples where a human has drawn the lines. But because the boundaries between lung segments are invisible, humans can't draw them accurately either. Asking a computer to guess the lines without a guide is like asking someone to draw the borders of countries on a blank piece of paper without a map.

2. The Solution: Following the "Highways"

The authors realized that while the neighborhoods are blurry, the roads and utilities inside them are very clear.

The Bronchovascular Tree: This is the network of airways (where air goes) and blood vessels (where blood goes).
The Analogy: Think of the lung segments as houses. The airways and arteries are the power lines and water pipes running into each house. Even if you can't see the walls of the houses, you can clearly see the pipes. If you know which pipe goes to which house, you can figure out where the house boundaries are.

3. The Two-Stage Strategy: "Build the Roads First"

The AI uses a two-step process, like a construction crew:

Stage 1: Map the Infrastructure.
First, the AI ignores the blurry neighborhoods and focuses entirely on drawing the "roads" (the arteries and airways). It creates a perfect map of the pipes and wires.
Stage 2: Build the Neighborhoods.
Now, the AI takes that perfect map of pipes and uses it as a guide. It says, "Okay, since I know exactly where the pipes are, I can now draw the walls of the houses around them." This ensures that every house (segment) contains the correct pipes, just like real anatomy requires.

4. The "Hierarchy" Rule: The Boss and the Subordinates

The AI is taught a strict rule of order, like a corporate hierarchy:

The Boss (The Lobe): The lung is divided into 5 big sections called lobes. The AI is told, "You must stay inside your boss's office."
The Employees (The Segments): Inside each lobe, there are smaller segments. The AI is told, "You must stay inside your lobe, and you must make sure you cover all the pipes assigned to you."

This prevents the AI from getting confused and mixing up segments from different lobes.

5. Smoothing the Edges: The "Jagged Edge" Fix

Sometimes, even with the right guide, the AI's drawing looks jagged and bumpy, like a pixelated video game. In real life, lung boundaries are smooth curves.

The Fix: The authors added a special "consistency loss" (a penalty system). If the AI draws a jagged line, it gets a "frown." If it draws a smooth, flowing line, it gets a "thumbs up." This forces the AI to smooth out the borders, making the final map look natural and medically useful.

6. The "Hole" Test: Checking for Quality

How do they know the map is good? They invented a new test called #holes.

The Analogy: Imagine the lung segments are islands. If the map is bad, the islands might have holes in the middle (like a donut) or be broken into tiny, scattered pieces.
The Metric: They count the number of "holes" or broken pieces in the map. The fewer holes, the smoother and more accurate the segmentation. Their method reduced the number of holes by over 90% compared to other methods.

The Bottom Line

This paper presents a smart way to teach AI to map the lungs. Instead of trying to guess invisible walls, it:

Maps the clear pipes first.
Uses those pipes to build the walls.
Enforces strict rules about which pipe belongs to which room.
Polishes the edges until they are smooth.

The result is a highly accurate, smooth map of the lung's 18 segments, which helps surgeons plan operations with much greater confidence and safety.

1. Problem Statement

Challenge: Accurate segmentation of pulmonary segments (18 distinct regions in the lungs) is critical for cancer localization and surgical planning. However, unlike lung lobes (5 regions), pulmonary segments lack distinct physical boundaries in CT images, making pixel-wise annotation extremely laborious and prone to error.
Limitations of Existing Methods:

Fully Supervised Learning (FSL): Requires precise pixel-level masks for all 18 segments, which is difficult to obtain due to indistinguishable boundaries.
Existing Weakly Supervised Learning (WSL): Previous WSL methods often rely on pseudo-labels or bronchovascular features but fail to strictly adhere to clinical anatomical definitions. Consequently, they often produce segments that do not fully encompass the associated bronchovascular structures (arteries and airways) or fail to respect lobe boundaries.
Patch Inference Issues: Standard 3D segmentation using small patches often breaks the structural integrity of the bronchovascular tree, leading to discontinuous or incorrect segment boundaries when branches from different segments are close together.

2. Methodology: Anatomy-Hierarchy Supervised Learning (AHSL)

The authors propose AHSL, a weakly supervised framework that leverages the precise clinical definition of pulmonary segments to guide the model without requiring full pixel-wise segment masks.

A. Clinical Anatomical Definition

The method is grounded in three clinical criteria for a valid pulmonary segment:

Lobe Containment: The segment must be entirely within its corresponding pulmonary lobe.
Vascular Inclusion: The segment must encompass its associated pulmonary artery and airway.
Boundary Definition: Intersegmental veins act as boundaries, while intrasegmental veins are included within the segment.

B. Two-Stage Segmentation Strategy

To address the loss of structural information during patch inference, AHSL employs a two-stage approach:

Stage 1 (Bronchovascular Segmentation): A U-Net backbone (based on nnUNet) segments the arteries and airways using Fully Supervised Learning (FSL) with Dice and Cross-Entropy losses. Veins are excluded in this stage due to their ambiguous boundary nature.
Stage 2 (Segment Segmentation): The output mask from Stage 1 is concatenated with the original CT image and fed into a second network. This provides priori knowledge of the bronchovascular tree, guiding the model to assign the correct segment label to the surrounding tissue.

C. Anatomy-Hierarchy Supervised Learning (Loss Function)

Instead of direct pixel-wise supervision for segments, the model uses a hierarchical supervision strategy based on two principles:

Segment-Level Supervision (Direct): The model is directly supervised on the Bronchovascular (BV) region. The loss ensures that voxels within the BV are correctly classified and that surrounding voxels are encouraged to be included in the corresponding segment.
- Loss: $Loss_{direct} = Recall_{BV} + CE_{BV}$
Lobe-Level Supervision (Indirect): The model is supervised on the Lobe (L) region. The probability of a voxel belonging to a lobe is derived from the maximum probability of the segments contained within that lobe. This ensures segments stay within their parent lobe and prevents over-segmentation.
- Loss: $Loss_{indirect} = Dice_{L} + CE_{L}$

D. Boundary Consistency Loss

To address irregular boundaries caused by insufficient constraints, a Consistency Loss is introduced. It calculates the L1-norm of the Laplacian of the probability map, penalizing large differences between a voxel's probability and its six neighbors. This enforces local smoothness in the output masks.

Total Loss: $Loss = Loss_{direct} + \lambda_1 Loss_{indirect} + \lambda_2 Loss_{consistence}$

E. Evaluation Metric: #holes

A new metric, #holes, is proposed to quantitatively measure the topological smoothness of the segmentation. It counts the number of "holes" (enclosed regions) in axial, sagittal, and coronal planes. A lower #holes score indicates smoother, more anatomically consistent boundaries.

3. Key Contributions

AHSL Framework: A novel weakly supervised method that strictly adheres to clinical anatomical definitions by utilizing lobe-level and segment-level (bronchovascular) labels.
Two-Stage Strategy: A mechanism to incorporate bronchovascular prior information, solving the issue of structural discontinuity in patch-based inference.
Consistency Loss: A regularization term specifically designed to enhance the geometric smoothness of pulmonary segment boundaries.
#holes Metric: A novel evaluation metric to quantify boundary smoothness and topological complexity.

4. Experimental Results

The method was evaluated on a private dataset of 200 CT scans (104 CTPA, 96 NCCT), split into 159 training and 41 testing samples.

Quantitative Performance (vs. IPGN):
- Artery Dice: AHSL achieved ~0.92 (CTPA/NCCT) compared to IPGN's ~0.70, representing a >30% improvement.
- Airway Dice: AHSL achieved ~0.92 compared to IPGN's ~0.51-0.55, representing a 70-80% improvement.
- Smoothness (#holes): AHSL drastically reduced the number of holes from 180-200 (IPGN) to **1.2-1.5**, indicating significantly smoother boundaries.
Ablation Study:
- Removing the two-stage strategy reduced Dice scores and increased #holes significantly.
- Removing the consistency loss maintained Dice scores but resulted in a 93-94% increase in #holes, confirming its necessity for boundary smoothness.
External Validation:
- Tested on 21 multi-center datasets from different equipment.
- Two radiologists scored the results on a 5-point scale, yielding mean scores of 3.67 and 3.71 (where 3 is "acceptable" and 5 is "perfect"), demonstrating clinical robustness.

5. Significance

This paper addresses a critical bottleneck in lung image analysis: the difficulty of annotating pulmonary segments. By shifting from pixel-wise supervision to anatomy-guided weak supervision, the authors enable accurate segmentation without the prohibitive cost of manual labeling. The method ensures that the resulting segments are not only geometrically accurate but also clinically valid (containing the correct vessels and staying within lobes). The proposed consistency loss and #holes metric offer new tools for improving and evaluating the topological quality of medical segmentation, making the technology more viable for surgical planning and cancer localization.