Progressive Curriculum Learning with Scale-Enhanced U-Net for Continuous Airway Segmentation

TL;DR

The paper tackles the challenge of continuous airway segmentation in chest CTs, which is hindered by intra-class imbalance and blurred distal airways. It introduces a three-stage progressive curriculum learning pipeline plus a Scale-Enhanced U-Net with Detail Information Enhancers, along with a robust airway tree parsing and hierarchical evaluation to measure continuity across airway scales. Key contributions include the three-stage sampling strategy with a dynamic scheduler, the Adaptive Topology-Responsive Loss to emphasize centerline continuity, and the SE-UNet architecture that preserves multi-scale details. Experimental results on the ATM’22 dataset and an in-house dataset demonstrate improved small-airway detection and airway tree completeness, outperforming several state-of-the-art methods and offering more reliable segmentation for clinical planning and navigation.

Abstract

Continuous and accurate segmentation of airways in chest CT images is essential for preoperative planning and real-time bronchoscopy navigation. Despite advances in deep learning for medical image segmentation, maintaining airway continuity remains a challenge, particularly due to intra-class imbalance between large and small branches and blurred CT scan details. To address these challenges, we propose a progressive curriculum learning pipeline and a Scale-Enhanced U-Net (SE-UNet) to enhance segmentation continuity. Specifically, our progressive curriculum learning pipeline consists of three stages: extracting main airways, identifying small airways, and repairing discontinuities. The cropping sampling strategy in each stage reduces feature interference between airways of different scales, effectively addressing the challenge of intra-class imbalance. In the third training stage, we present an Adaptive Topology-Responsive Loss (ATRL) to guide the network to focus on airway continuity. The progressive training pipeline shares the same SE-UNet, integrating multi-scale inputs and Detail Information Enhancers (DIEs) to enhance information flow and effectively capture the intricate details of small airways. Additionally, we propose a robust airway tree parsing method and hierarchical evaluation metrics to provide more clinically relevant and precise analysis. Experiments on both in-house and public datasets demonstrate that our method outperforms existing approaches, significantly improving the accuracy of small airways and the completeness of the airway tree. The code will be released upon publication.

Figures (7)

• Figure 1: Example of airway discontinuity. (a) Green mask: ground truth. (b) Blue mask: prediction based on Zhang et al. zhang2023towards. (c) Red mask: prediction based on our method. Blurry airway walls lead to discontinuities, which were not effectively addressed by zhang2023towards (yellow box).
• Figure 2: Overview of the proposed progressive curriculum learning pipeline for airway segmentation. In “Stage 1”, the network is trained utilizing the Dice Loss milletari2016v. The epoch with the highest accuracy from the “Stage 1” forms the basis for the hard-mining strategy employed in the “Stage 2”. This stage focuses on learning to detect the challenging airways with GUL zheng2021alleviating. Building upon the predictions with the highest recall from “Stage 2”, the “Stage 3” introduces ATRL to enhance the length of the airway tree. Ultimately, the optimal model is selected for segmentation testing.
• Figure 3: The SE-UNet architecture in proposed method. “$\text{DIE}_1$$8/16/32$” refers to the output channels of the three ConvBlocks in DIE1 are 8, 16 and 32. The detailed structure of DIE is shown in the orange box.
• Figure 4: Explanation of identifying airway breakages. Here shows an airway skeleton segment, with purple points representing successful predictions and white points indicating failures. A 3×3×3 all-ones convolution kernel is applied to the white points to count their neighbors. If no point on a challenging airway segment has a single neighbor, it is identified as a breakage.
• Figure 5: A qualitative example illustrating the topological analysis process of airway tree.