Back to all papers

KD-SqueezeNet: an efficient deep learning strategy for the multi-task diagnosis of neonatal lung diseases.

February 20, 2026pubmed logopapers

Authors

Li J,Pan R,Tian F,Pan P,Yan Z

Affiliations (7)

  • The Department of Radiology, Second Affiliated Hospital and Yuying Childrens Hospital of Wenzhou Medical University, Xueyuan West Road, Wenzhou, 325027, China.
  • The Department of Radiology, First Affiliated Hospital of Wenzhou Medical University, Wenzhou, China.
  • Hangzhou Normal University, Hangzhou, China.
  • Sichuan Agricultural University, Yucheng, China.
  • Wenzhou Key Laboratory of Structural and Functional Imaging, Wenzhou, China.
  • The Department of Radiology, Second Affiliated Hospital and Yuying Childrens Hospital of Wenzhou Medical University, Xueyuan West Road, Wenzhou, 325027, China. [email protected].
  • Wenzhou Key Laboratory of Structural and Functional Imaging, Wenzhou, China. [email protected].

Abstract

The integration of deep learning in medical imaging has reached proficiency levels akin to expert clinicians, particularly in tasks requiring precise image categorization. This study developed KD-SqueezeNet, a lightweight deep learning model, to classify neonatal lung diseases via chest radiographs, aiming to enhance diagnostic accuracy and efficiency. Retrospective analysis included 2,089 neonates with clinical and imaging records. Chest radiographs were categorized into five groups: bronchopulmonary dysplasia (group 1, n=205), pneumonia (group 2, n=505), pneumothorax (group 3, n=201), respiratory distress syndrome (group 4, n=629), and normal (group 5, n=549). Data were divided into training, testing, and validation sets with an 8:1:1 ratio. Performance metrics included accuracy (Ac), precision (Pr), recall (Rc), F1 score (F1), parameter count, and area under the receiver operating characteristic curve (AUROC). KD-SqueezeNet, an interpretable model integrating knowledge distillation, outperformed EfficientNet, GhostNet, InceptionNet, RegNet, and Vision Transformer. For binary classification (healthy/diseased), it achieved Ac=0.93, Pr=0.93, Rc=0.93, F1=0.93, and AUROC=0.97 with only 723,522 parameters. In four-class classification (group 1/group 2/group 3/group 4), it attained Ac=0.86, Pr=0.86, Rc=0.86, and F1=0.86 (724,548 parameters), with class-specific AUROCs: group 1 (0.97), group 2 (0.94), group 3 (0.96), group 4 (0.97). KD-SqueezeNet not only excels in accuracy and stability but also demonstrates efficient utilization of computational resources, making it suitable for rapid diagnosis and deployment in practical applications. It holds significant clinical value in terms of saving time and server space, providing auxiliary diagnostics, supporting clinical decision-making, and improving patient outcomes in large-scale screening contexts.

Topics

Journal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAI Slice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.