Abstract
This paper investigates six deep-learning models for segmenting the short axis of the abdominal aorta in point-of-care ultrasound (POCUS) images. A total of 724 images from 175 adult patients were collected from various perspectives in a remote area. The dataset was split into three sets: 510 images (146 patients) for training, 74 images (from 17 patients) for validation, and 140 images (18 patients) for testing. The six deep learning models utilized in this research were UNet, Attention UNet, Res-UNet, Res-Attention UNet, YOLO (You Look Only Once), and a combination of YOLO with the Segment Anything model (SAM). The models were trained on the training dataset and during training hyperparameters were selected based on the models' performance on validation set. Res-Attention UNet achieved the highest Dice Similarity Score (DSC) on the training Dataset, (0.907) and the validation dataset (0.858). However, YOLO stood out as the top model with a DSC of 0.857 on the testing dataset, showing a reliable and effective segmentation performance. Furthermore, the models were additionally evaluated on an independent dataset of 375 images from 375 patients with mean DSC were YOLO + SAM: 0.763, YOLO: 0.759, UNet: 0.666, ResUNet: 0.618, Attention UNet: 0.687, and Res Att. UNet:0.693. When trained with 50% of the data, YOLO models significantly outperform UNet models, with the addition of SAM to YOLO (YOLO + SAM) only marginally affecting performance. The paper also introduces a user-friendly web-based Aorta segmentation tool, aiming to enhance reader’s experience by performing hands-on experiments on YOLOv8 model.