Deep Learning for Ultrasound-Based Auxiliary Diagnosis of Emergency Ascites.
Authors
Affiliations (6)
Affiliations (6)
- Cancer Centre, Department of Ultrasound Medicine, Zhejiang Provincial People's Hospital, Affiliated People's Hospital, Hangzhou Medical College, Hangzhou, Zhejiang, China.
- School of Mathematical Sciences, Zhejiang University, Hangzhou, Zhejiang, China.
- Cancer Centre, Department of Ultrasound Medicine, Zhejiang Provincial People's Hospital, Affiliated People's Hospital, Hangzhou Medical College, Hangzhou, Zhejiang, China; School of Mathematical Sciences, Zhejiang University, Hangzhou, Zhejiang, China.
- Department of Ultrasound Medicine, Renmin Hospital of Wuhan University, Wuhan, Hubei, China.
- Department of Thoracic Surgery, The First Affiliated Hospital of Jinzhou Medical University, Jinzhou, Liaoning, China.
- Cancer Centre, Department of Ultrasound Medicine, Zhejiang Provincial People's Hospital, Affiliated People's Hospital, Hangzhou Medical College, Hangzhou, Zhejiang, China. Electronic address: [email protected].
Abstract
The Focused Assessment with Sonography for Trauma (FAST) enables rapid detection of free intraperitoneal fluid, facilitating timely management of internal hemorrhage. This study aims to develop a Transformer-based model for automated detection on FAST images and evaluate its feasibility in assisting non-specialist operators. Between June 2019 and June 2024, 1829 FAST-positive images demonstrating free intraperitoneal fluid and 303 FAST-negative images without fluid were retrospectively collected from Zhejiang Provincial People's Hospital. A Transformer-based model integrating segmentation and classification modules was developed and internally validated using five-fold cross-validation. External validation was performed on 848 images (424 positive/424 negative) from Hubei Provincial People's Hospital. Three operators with varying expertise-a junior sonographer, a clinician, and a non-clinical operator-evaluated all external images before and after model assistance to compare segmentation performance. Five-fold cross-validation yielded the following segmentation metrics: mean intersection over union (IoU) 0.671 ± 0.009, Dice coefficient 0.799 ± 0.013, and pixel accuracy (PA) 0.809 ± 0.010. Classification performance showed a mean accuracy of 0.922 ± 0.013, sensitivity of 0.938 ± 0.015, and specificity of 0.905 ± 0.015. External validation demonstrated accuracy 0.883, sensitivity 0.901, specificity 0.837, AUC 0.871, with segmentation IoU 0.683, Dice coefficient 0.812, and PA 0.781. The model performed comparably to junior Sonographer and outperformed non-specialists. After model assistance, all groups improved and inter-group differences disappeared (all p > 0.05). The Transformer model delivers diagnostic performance comparable to junior Sonographer for automated free intraperitoneal-fluid detection in FAST examinations and significantly improves detection accuracy among non-specialist operators.