RSNA 2018 Pneumonia 3 Level Severity Split Dataset

This dataset presents a severity-oriented reformulation of the RSNA 2018 Pneumonia Detection Challenge chest X-ray collection, designed to support research in pneumonia grading, severity-aware classification, and explainable medical AI. Starting from the publicly available RSNA 2018 dataset and its expert bounding-box annotations, we derive structured severity labels by quantifying lesion burden at image level. For each pneumonia-positive study, the total disease area is computed as the sum of all annotated bounding-box areas, allowing radiographic burden to be represented as a continuous quantitative measure. Based on this burden signal, pneumonia-positive images are stratified into three severity categories: Severity 1 (low/mild), Severity 2 (moderate), and Severity 3 (severe). The resulting dataset contains 8,851 normal chest X-ray images and 6,012 pneumonia-positive images with expert bounding-box annotations. The positive cohort is further divided into three severity classes comprising 1,999 Severity 1 images, 2,006 Severity 2 images, and 2,007 Severity 3 images. The corresponding thresholds are defined from bounding-box extent as follows: Severity 1 for total lesion area below 60,455 px(^2), Severity 2 for total lesion area between 60,455 and 139,133 px(^2), and Severity 3 for total lesion area greater than or equal to 139,133 px(^2). In total, the dataset preserves both the original normal class and a balanced three-level severity structure for pneumonia-positive studies. This resource is intended for researchers working on chest X-ray analysis, computer-aided diagnosis, severity-aware deep learning, and interpretable AI systems for medical imaging. By bridging lesion localization metadata and radiographic burden stratification, the dataset extends the utility of the RSNA 2018 benchmark beyond binary pneumonia detection and localization, enabling new investigations into severity grading, triage-oriented modeling, and clinically meaningful chest X-ray classification.