RSNA Pneumonia Detection

Overview

Detect and localize lung opacities suggestive of pneumonia in chest radiographs. This object detection challenge uses DICOM images from the RSNA and STR, requiring both classification and bounding box prediction.

Data Description

The dataset contains frontal chest radiographs in DICOM format:

• patientId — Unique patient identifier
• x, y, width, height — Bounding box coordinates for opacity regions (training only)
• Target — Binary label (1 = opacity present, 0 = normal)

Images vary in resolution and bit depth. Multiple bounding boxes may exist per image when several opacity regions are present. Images labeled as normal have no bounding box annotations.

Additional metadata from the DICOM headers includes patient age, sex, and view position.

Evaluation

Submissions are evaluated using mean Average Precision at IoU thresholds from 0.4 to 0.75 in steps of 0.05.

For each IoU threshold, a predicted box is a true positive if it overlaps a ground truth box above the threshold. mAP is computed as the mean of per-class AP values across all thresholds.

Submission Format

Submit a CSV with the following columns:

For normal predictions (no opacity), leave PredictionString empty.