Pneumonia represents a respiratory disease with high incidence and mortality rates in childhood. The accurate segmentation of lung computed tomography (CT) images plays a crucial role in early diagnosis and treatment planning. The manual labeling of infected regions, however, is time-intensive and burdensome, significantly increasing radiologists’ workload. Therefore, developing efficient automatic segmentation methods holds substantial practical significance in alleviating medical resource constraints. Current predominant medical image segmentation approaches primarily utilize U-shaped architecture, known for its semantic modeling capabilities. However, the encoder-decoder structure inherently requires multiple down-sampling operations, resulting in the loss of critical spatial structure information and compromising segmentation accuracy. Furthermore, infected regions in childhood pneumonia CT images typically present with scattered and fragmented multifocal distribution patterns, demanding enhanced model capabilities for capturing long-distance dependencies and maintaining overall structural coherence. While Transformer-based segmentation networks have demonstrated strong performance in global modeling recently, their limited local spatial priors and patch size constraints often lead to inadequate local detail segmentation. Additionally, normal anatomical structures such as the lung hilum exhibit morphological similarities to infected regions, necessitating superior network anti-interference capabilities. To address these challenges, this study proposes a U-Net-based difference aware guided boundary Transformer segmentation network.

DBTU-Net aims to enhance both local structural detail modeling and global semantic dependency modeling capabilities, enabling accurate segmentation of fragmented and scattered multi-focal regions. The network architecture builds upon the classical U-shaped structure, incorporating three key components: gated channel Transformer (GCT), difference aware fusion (DAF), and boundary Transformer (BT), as shown in Fig. 1. During the feature extraction phase, multi-scale semantic information is progressively extracted through multilayer convolution and down-sampling operations. To enhance the network’s context modeling capability, the GCT module (Fig. 2) is integrated into each encoder layer. This module dynamically models channel dependencies to adaptively adjust the importance distribution across different semantic channels, thereby strengthening global information perception. The DAF module is implemented in the skip connection path, explicitly enhancing spatial structure by computing difference information between adjacent encoder layer features. This mechanism mitigates spatial detail loss from down-sampling while providing the decoder with comprehensive structural priors, improving the model’s capacity to recognize small lesion regions. At the network’s bottleneck layer, the BT module (Fig. 3) further enhances global modeling capability. This module utilizes encoder multi-scale disparity maps for guidance, establishing potential connections between distant lesion regions through Transformer architecture, improving distant lesion recognition, and refining boundary segmentation while maintaining global consistency. The decoder ultimately produces a high-quality segmentation map through up-sampling operations.

Experimental analyses are conducted on a private childhood pneumonia CT dataset (Child-P) and two public COVID-19 CT datasets (COVID, MosMed) to validate DBTU-Net’s effectiveness. Eight ablation schemes were designed to evaluate the performance of three key modules: GCT, DAF, and BT. The results demonstrate that each module enhances network segmentation performance (Table 2), with the DAF module showing notable improvements of 8.21 percentage points, 12.34 percentage points, and 13.94 percentage points for Dice similarly coefficient (Dice), Jaccard index (JI), and sensitivity (SE) metrics, respectively, confirming its effectiveness in enhancing spatial detail expression, preserving structural information, and improving lesion integrity. Module combination experiments further validate the DAF module’s importance. Without DAF, retaining only GCT and BT leads to performance decreases of 1.09 percentage points and 0.99 percentage points in JI and SE metrics compared to the full model. The DAF and BT combination achieves 25.24 pixel in Hausdorff distance (HD) metrics while maintaining high comprehensive performance, demonstrating their synergistic effect on boundary detail extraction. In comparison experiments, DBTU-Net achieves optimal results across all five metrics on the Child-P dataset (Table 3), with Dice and JI reaching 89.61% and 81.17%, representing improvements of 8.17 percentage points and 12.48 percentage points over the baseline network, surpassing the suboptimal CASCADE network. Visualization results indicate DBTU-Net’s superior sensitivity in identifying scattered and tiny lesions, reducing missed segmentation and over-segmentation instances (Fig. 6). The decreased HD metrics demonstrate the model’s advantage in lesion boundary modeling, validating the effectiveness of cross-scale difference perception and boundary modeling mechanisms. On the COVID dataset, DBTU-Net leads in all five metrics, with JI, SE, and Matthews correlation coefficient (MCC) reaching 69.66%, 77.02%, and 81.04%, significantly outperforming U-Net++ in balancing lesion pixel identification and background differentiation (Table 4). On the MosMed dataset, despite slightly lower SE than TransDeepLab, DBTU-Net achieves optimal results in Dice, JI, and MCC metrics (Table 5), demonstrating robust lesion structure modeling. Visualization results show DBTU-Net’s advantage in reducing mis-segmentation and false positives, attributed to its contextual modeling mechanism integrating GCT and BT (Fig. 7). Multiple local detail visualizations confirm DBTU-Net maintains consistent segmentation in regions with blurred edges or irregular lesion morphology, validating its local segmentation accuracy and robustness under complex structures (Fig. 8).

This research focuses on childhood pneumonia segmentation. DBTU-Net addresses the limitations of traditional U-shaped networks in segmenting small, fragmented, and complex lesions due to spatial information loss and limited global feature extraction. The network enhances spatial structure expression by utilizing the DAF module to mine differential features between layers, while incorporating the BT module to guide high-level semantic information for boundary enhancement. This combination improves the modeling capability for distant lesions and local boundary segmentation accuracy, reducing mis-segmentation in complex lesion regions. Experimental results on the private childhood pneumonia dataset demonstrate DBTU-Net’s superior performance compared to existing mainstream methods across multiple evaluation metrics. Additionally, its strong performance on the public COVID-19 dataset validates the method’s generalization capabilities.