Researchers from the First Affiliated Hospital of Guangzhou Medical University and Hengqin Hospital in China have published a publicly available, high-quality annotated dataset of 2,756 ultrasound images of pulmonary tumors. The dataset — PU2756 — and background research are detailed in the paper, “A Pulmonary Ultrasound Dataset for Tumor Segmentation and Classification,” published in Nature.
According to the authors, this is the first openly accessible resource of its kind and caliber to help health care professionals discern malignant and benign tumors. Expert sonographers annotated each image in the dataset, which is available on figshare, for tumor segmentation and classification, and all diagnostic results were confirmed through biopsy.
The dataset focuses on peripheral pulmonary lesions (PPLs) of 2,756 unique patients, and images were marked with patient-level metadata, benign-malignant labels and pixel-level lesion masks. The researchers used stratified cross-validation to partition the data into five folds. This assured a balance between training and testing sets in the distributions of benign and malignant cases, they said.
Accurate tumor demarcation is vital for clinical decision-making and improved patient outcomes, and ultrasound remains the most noninvasive, cost-effective, accessible method of diagnostic evaluation, researchers noted.
The free resource serves as an important benchmark for pulmonary ultrasound and “may facilitate future research in AI-assisted pulmonary ultrasound analysis,” the authors wrote.
