Artificial intelligence models for point-of-care ultrasound diagnostics in dogs.

Abstract

INTRODUCTION: Point-of-care ultrasound (POCUS) for the purpose of Focused Assessment with Sonography for Trauma (FAST) is an essential diagnostic tool for triage in canine patients, but its accuracy is highly operator-dependent. Artificial intelligence (AI) offers a potential solution for improving diagnostic capability by providing real-time, automated interpretation of ultrasound images, particularly in resource-limited or pre-hospital settings. This study evaluated the feasibility and diagnostic performance of deep learning models for detecting life-threatening effusions and pneumothorax (PTX) in dogs. METHODS: Five healthy military working dogs (MWDs) and twenty client-owned dogs (22-55 kg) were prospectively enrolled. MWDs were negative for injury for baseline data capture. Client-owned dogs with confirmed abdominal, pleural, pericardial effusion, or PTX were imaged using POCUS. Ultrasound clips were reviewed for quality, curated by experts, converted to image frames from videos, and used to train, optimize, and evaluate different convolutional neural network (CNN) architectures at all FAST scan sites. RESULTS: Models were developed for each scan site with varied performance. Diaphragmatico-hepatic scan site models achieved excellent performance (recall 98%, accuracy 97%) while the pericardial models (recall 87%, accuracy 85%) and chest tube site models (recall 81%, accuracy 88%) demonstrated good performance. The spleno-renal/hepato-renal models (recall 83%, accuracy 78%) and cysto-colic models (recall 84%, accuracy 77%) achieved fair performance. Model prediction overlays confirmed that the models for each site focused on clinically relevant regions during predictions. DISCUSSION: Deep learning models can accurately detect effusion and PTX in canines using POCUS, with variable performance at individual sites. Limitations included small sample sizes, inclusion of only blunt trauma and non-traumatic pathology, class imbalances, and variability in the volume and location of effusion on presentation. Expanding the training datasets and refining pre-training strategies may enhance performance. These findings support the feasibility of AI-assisted ultrasound to augment triage and pre-hospital decision-making in veterinary emergency care.