Improved U-Net for Automated Segmentation of Lung Ultrasound Findings and Assistance with Lesion Assessment in Children with Community-Acquired Pneumonia
Year: 2026
Type: article
Abstract: Objectives: To evaluate an improved U-Net for automatic segmentation of lung ultrasound (LUS) findings in pediatric community-acquired pneumonia (CAP) and its effect on sonographers’ contouring performance and efficiency. Methods: This prospective study included 1994 images from 441 children: 1019 with consolidation, 767 with B-lines, and 208 with pleural effusion. The patient-level development and internal held-out set comprised 353 children (1575 images) and 88 children (419 images), respectively. The model used a ResNet34 encoder, dense atrous convolution, residual multi-kernel pooling, and weighted binary cross-entropy losses for multiscale segmentation deep supervision. We evaluated six model configurations by five-fold cross-validation and on an internal held-out test set. The complete model was also compared with sonographers, and two less-experienced readers repeated contouring with model assistance. Image-level paired differences were analyzed using 50,000 patient-cluster bootstrap replicates and within-family Holm adjustment. Results: Development-set Dice similarity coefficients (DSCs) for consolidation, B-lines, and pleural effusion were 0.8411 ± 0.0107, 0.7698 ± 0.0142, and 0.7730 ± 0.0187, respectively; internal held-out set DSCs were 0.8159 ± 0.0134, 0.7467 ± 0.0178, and 0.7498 ± 0.0234. The improved model had higher DSC, intersection over union, and boundary F1 scores and lower 95th-percentile Hausdorff distances than baseline U-Net for all three findings. It exceeded both less-experienced sonographers but remained below senior sonographers. With assistance, both less-experienced readers improved DSC across all findings and exceeded the standalone model in DSC, and reduced mean per-image processing time by 38.3% and 35.8%, respectively. Conclusions: The improved U-Net segmented consolidation, B-lines, and pleural effusion on LUS images from children with CAP and outperformed baseline U-Net on segmentation metrics. Model assistance improved lesion contouring performance in both less-experienced sonographers. These findings highlight the potential of the model as an AI-assisted tool to facilitate lesion delineation and improve the efficiency of LUS assessment in pediatric CAP and may help reduce reliance on operator experience during contouring.
Source: Diagnostics
Authors: Tong Su(Conceptualization, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft), Sipeng Tang(Conceptualization, Formal analysis, Methodology, Software, Visualization, Writing – review & editing), Yang Li(Formal analysis, Investigation, Visualization, Writing – review & editing), Lei Zhong(Formal analysis, Investigation, Supervision, Writing – review & editing), Yajun Wang(Methodology, Supervision, Visualization, Writing – review & editing), Yingtao Zhang(Conceptualization, Methodology, Project administration, Software, Supervision, Writing – review & editing), Binbin Guo(Conceptualization, Data curation, Methodology, Project administration, Supervision, Validation, Writing – review & editing)
Institutions: Harbin Medical University, Harbin Institute of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,137
Funder: Harbin Medical University
OpenAlex ID: W7235439914
