Application of artificial intelligence-assisted ultrasound in the segmentation of splenic trauma

Abstract

To develop an artificial intelligence(AI)-assisted ultrasound segmentation model for splenic trauma and to evaluate its clinical utility for trauma localization and contour delineation. (1) We assembled an animal ultrasound dataset comprising 4,280 images and a clinical dataset comprising 117 cases (318 human ultrasound images) of splenic trauma. Among the clinical images, 123 were used as a training set to adapt the baseline animal-trained model. The remaining clinical images were used for independent evaluation: 77 splenic trauma images collected at the Chinese People’s Liberation Army General Hospital served as the internal test set, and 118 images collected at the Beijing Chaoyang Emergency Rescue Center served as the external test set. (2) Lesion segmentation and contour delineation were independently performed by three senior ultrasound physicians (each with at least 15 years of experience). The physicians manually outlined splenic trauma lesions as precisely as possible, using contrast-enhanced ultrasound (CEUS) and/or contrast-enhanced computed tomography (CT) as the reference standard when available. (3) We applied transfer learning to fine-tune the animal-trained model using the human splenic trauma dataset. For performance comparison, three ultrasound physicians with different experience levels (a junior physician with two years of standardized training, an intermediate physician with five years of clinical ultrasound experience, and a senior physician with ten years of clinical ultrasound experience) independently assessed splenic trauma, and their results were compared with the artificial intelligence outputs. (1) The Dice coefficients for AI and ultrasound physicians of different seniority in Internal Test Set and External Test Set were as follows: Internal Test Set: 0.71, 0.79, 0.79, and 0.79; External Test Set: 0.74, 0.76, 0.77, and 0.79. (2) Across different American Association for the Surgery of Trauma (AAST)grades, the AI Dice coefficients were higher than those of junior and intermediate physicians for AAST grade II (Internal Test Set: 0.71 vs. 0.68 and 0.67; External Test Set: 0.76 vs. 0.73 and 0.74) and AAST grade III (Internal Test Set: 0.75 vs. 0.71 and 0.72; External Test Set: 0.76 vs. 0.70 and 0.72) and were closer to the performance of senior physicians. (3) For stratified analyses by AAST grade, AI and senior physicians showed greater reproducibility between Internal Test Set and External Test Set than junior and intermediate physicians. AI demonstrated strong performance for ultrasound-based segmentation of splenic trauma and maintained stable diagnostic consistency across test sets and injury grades.