Efficient Deep Learning Models Revolutionize Doctor’s Training for Point-of-Care Ultrasound

Point-of-care ultrasound (PoCUS) is a valuable diagnostic tool for pericardial effusion (PCE). However, the time constraints for trainees and experts pose significant barriers to PoCUS learning. This study aims to develop a deep learning (DL) model for detecting and localizing PCE and to investigate the learning efficacy of using this model as an adjunct in PoCUS training. A total of 101 patients with moderate/large PCE and 104 controls without PCE were included, and the images were extracted from the ultrasound (US) clips.We applied preprocessing techniques, including standardized image sizes and background removal, to reduce interference, and post-processing techniques, including adding filters to refine small effusion regions. We developed three DL models based on U-Net, Res-UNet, and UNet++ and compared their performance. Additionally, 14 emergency medicine residents were recruited to complete classification and segmentation tasks on 10% of randomly selected US images. Personalized feedback from the best-performing DL model was provided. Three months later, the residents annotated another set of images. Their learning performance was also evaluated. The UNet++ algorithm surpassed the other two, attaining an impressive sensitivity of 96%, specificity of 97%, area under the curve (AUC) of 98%, intersection over union (IOU) of 81%, and minimal latency. The overall sensitivity increased by 4% in the classification task after training with the UNet++ model, although there were no statistically significant differences in all evaluation metrics. The UNet++ model achieved a balance between high accuracy, IOU, and latency. Although there were no significant differences in the evaluation metrics after UNet++-assisted learning, the overall sensitivity increased by 4%, indicating an improved ability to recognize true positives and reduce false negatives. Our results demonstrated that AI could enhance the interpretation of rare PoCUS conditions and reduce the time demands on both trainees and teachers.