A dataset of lung ultrasound images for automated AI-based lung disease classification

Abstract

Lung ultrasound (LUS) is increasingly recognized as a valuable imaging modality for evaluating various pulmonary conditions. Despite its clinical utility, accurate interpretation of LUS remains challenging due to factors such as inter-operator variability, dependence on sonographer expertise, and inherently low signal-to-noise ratios. This article presents a curated benchmark dataset of labelled LUS images acquired in Uganda, intended to support the development of automated, AI-based diagnostic tools for lung disease classification. The dataset comprises 1062 labelled images collected from patients at Mulago National Referral Hospital and Kiruddu Referral Hospital by senior radiologists. The dataset is suitable for training and evaluating convolutional neural networkbased models and is expected to facilitate research in developing robust deep learning systems for pulmonary disease diagnosis using LUS