NLUS-VQA-VG: Enhancing Interpretability in Domain-Specific Med-VQA for Neonatal Lung Ultrasound Through Visual Grounding

Neonatal lung diseases present significant diagnostic challenges during the perinatal period. Lung ultrasound, a safe and bedside imaging technique, has garnered increasing attention, yet its interpretation remains dependent on the operator's expertise. We introduce NLUS-VQA-VG, a domain-specific Med-VQA model designed for neonatal lung ultrasound, which enhances diagnostic accuracy and interpretability through visual grounding methods. A dedicated image-text dataset was developed, incorporating color-coded bounding boxes for visual grounding annotations, and a three-stage reasoning framework was implemented to ensure precise image-text alignment. Built on the Qwen2.5-VL-7B foundation, the model is optimized via efficient parameter fine-tuning using LoRA. Qualitative and quantitative evaluations demonstrate that NLUS-VQA-VG outperforms baseline models in sign recognition, localization accuracy, and reduction of hallucinations. This study addresses a critical gap in the Med-VQA domain for neonatal lung ultrasound, aiding clinical decision-making and advancing precision medicine in neonatology.