A Pulmonary Ultrasound Dataset for Tumor Segmentation and Classification

Pulmonary tumor diagnosis with ultrasound is clinically valuable yet technically challenging. Although deep learning has shown promise in supporting sonographers, progress is hindered by the lack of publicly available, high-quality annotated datasets. Given the limited clinical adoption of pulmonary ultrasound for differentiating benign and malignant tumors, publicly available datasets for AI research in this area remain scarce. To address this gap, we present PU2756, an openly accessible dataset of 2,756 B-mode ultrasound images of pulmonary tumors, each annotated by expert sonographers for pulmonary tumor segmentation and classification tasks and pathologically confirmed through biopsy. We further provide statistical analysis and establish benchmark results using several machine learning and deep learning approaches. To our knowledge, PU2756 is among the first publicly available expert-annotated pulmonary ultrasound datasets specifically focused on peripheral pulmonary tumors to provide pixel-level lesion masks, pathologically confirmed benign-malignant labels, and patient-level metadata for 2,756 unique patients. This resource provides a publicly available benchmark dataset for pulmonary ultrasound image segmentation and benign-malignant classification and may facilitate future research in AI-assisted pulmonary ultrasound analysis.