Sonographic machine-assisted recognition and tracking of B-lines in dogs: the SMARTDOG study

Introduction Cardiogenic pulmonary edema (CPE) is a serious complication of heart failure in dogs, commonly characterized by excess fluid within the lung interstitium and alveoli. Point-of-care ultrasound (POCUS) allows for the prompt identification of pulmonary alterations through the presence of B-lines. However, interpretation remains subjective and operator dependent. Artificial intelligence (AI) may offer standardized, real-time analysis, but its application in veterinary medicine is largely unexplored. Objective To assess the performance of an AI-based ultrasound algorithm in detecting B-lines in dogs and to evaluate its agreement with manual quantification by experienced operators. Methods In this prospective study conducted at a single center, 40 dogs were enrolled: 20 with suspected CPE and 20 healthy controls. CPE suspicion was based on respiratory distress, a left atrium-to-aorta ratio (La:Ao) ≥1.6, >3 B-lines per field at thoracic POCUS, and clinical improvement following furosemide administration. Lung ultrasound was performed according to the Vet BLUE protocol. Cine loops were analyzed using the Butterfly Auto B-line Counter and reviewed independently by two POCUS-trained clinicians, each blinded to the AI results and to the other's evaluation. Results The AI algorithm failed to provide a B-line count in 14.2% of cineloops overall, with failures occurring in 11.8% of the suspected CPE group and 2.4% of the non-CPE group. Quantification failures were significantly more frequent in the suspected CPE group (OR 4.88; p < 0.0001). Intraclass correlation coefficients showed excellent agreement for B-line counts (ICC = 0.88) and strong concordance for pathological classification (>3 B-lines; ICC = 0.85) between operators and AI. AI accuracy compared to clinicians was 84 and 86%. Conclusion The AI algorithm demonstrated excellent agreement with experienced operators both for precise B-line counting and for the classification of pathological lung patterns. These findings support the potential of AI as a valuable decision-support tool for detecting clinically relevant cardiogenic pulmonary edema in veterinary critical care.