ttumyche/CheXStruct

Overview

The CheXStruct dataset, introduced in the paper CXReasonBench: A Benchmark for Evaluating Structured Diagnostic Reasoning in Chest X-rays, provides structured clinical information extracted from high-quality chest X-ray images using the CheXStruct pipeline, a fully automated framework for deriving diagnostic reasoning from chest X-rays.

The pipeline generates structured outputs for 12 diagnostic tasks, categorized as Radiological Findings and Image Quality Assessment, as well as 3 global filtering tasks to exclude non-frontal or corrupted images:

• Global Filtering

  • Excludes non-frontal or corrupted images by assessing the number of segmented anatomical structures, anatomical positioning (e.g., presence of substantial abdominal or pelvic anatomy), and post-processing artifacts (e.g., low contrast).

• Radiological Findings

  • Cardiomegaly, Mediastinal Widening, Carina Angle, Trachea Deviation, Aortic Knob Enlargement, Ascending Aorta Enlargement, Descending Aorta Enlargement, and Descending Aorta Tortuous

• Image Quality Assessment

  • Inclusion, Inspiration Level, Rotation, and Projection

---

Dataset Structure

The CheXStruct dataset is derived from three publicly available chest X-ray datasets and is organized into three main folders, each corresponding to one of the source datasets:

• nih_cxr14: Derived from the NIH Chest X-ray 14 dataset, available at https://www.kaggle.com/datasets/nih-chest-xrays/data.
• vindrcxr: Derived from the VinDr-CXR dataset, available at https://github.com/vinbigdata-medical/vindr-cxr.
• openi: Derived from the OpenI dataset, available at https://www.kaggle.com/datasets/raddar/chest-xrays-indiana-university.

The MIMIC-CXR-JPG dataset is not included in this repository due to requiring PhysioNet credential access but will be uploaded to PhysioNet soon!

Within each folder, there are 15 .csv files: 12 corresponding to the diagnostic tasks listed in the Overview, and 3 additional files (abdominal_xray.csv, mask_number.csv, window.csv) generated during a global filtering process to exclude non-frontal or corrupted chest X-ray images.

These files contain the structured outputs extracted by the CheXStruct pipeline for the respective task.

CSV File Structure

Each .csv file contains the following common columns across all diagnostic tasks:

• image_file: Unique identifier for the chest X-ray image in the respective dataset.
• viewposition: The chest X-ray view position, either PA (posteroanterior) or AP (anteroposterior). For datasets that do not support view position (e.g., OpenI), this is marked as N/A.
• label: Binary label (1/0) indicating the presence or absence of the specific radiological finding or the state of the image quality assessment, derived by the CheXStruct pipeline. The specific meaning of the label varies by task, as detailed below.

In addition to these common columns, each .csv file includes task-specific columns representing:

• Anatomical landmarks: Extracted landmarks derived from segmentation masks (e.g., cardiac end points, thoracic end points).
• Diagnostic measurements: Measurements calculated based on the anatomical landmarks (e.g., cardiac width, thoracic width).
• Diagnostic indices: Computed indices based on the diagnostic measurements (e.g., cardiothoracic ratio - CTR).

The specific column names vary by task, as follows:

Global Filtering

• mask_number.csv

  • label: 0 indicates an invalid image (e.g., too few structures detected, suggesting the image is not a chest X-ray, is corrupted or blank, or has anatomical structures obscured due to poor patient positioning or motion artifacts), 1 indicates a valid image.
  • mask_count: The count of valid segmentation masks, determined as the number of masks where the sum of pixel values is non-zero.

• abdomial_xray.csv

  • label: 0 indicates the presence of substantial abdominal or pelvic anatomy, 1 indicates a standard chest X-ray.
  • y_start: Y-coordinate of the top of the image
  • y_heart_max: Y-coordinate of the lowest point of the heart mask
  • y_end: Y-coordinate of the bottom of the image
  • distance_upper: Vertical distance from y_start to y_heart_max, representing the portion of the image above the heart
  • distance_lower: Vertical distance from y_heart_max to y_end, representing the portion of the image below the heart
  • ratio: Ratio of the vertical distances below and above the heart, calculated as distance_lower / distance_upper.

• window.csv:

  • label: 0 indicates the presence of post-processing artifacts, 1 indicates acceptable image quality.
  • gradient_mean: The mean gradient magnitude across the image, calculated as a metric of edge strength to detect post-processing artifacts.

Radiological Findings

• cardiomegaly.csv

  • label: 1 indicates the presence of cardiomegaly, 0 indicates its absence.
  • heart_xmin, heart_xmax: X-coordinates of the heart's left and right end points.
  • lung_xmin, lung_xmax: X-coordinates of the lung's left and right end points.
  • heart_width, lung_width: Widths of the heart and lungs, calculated from the coordinates heart_xmin, heart_xmax, lung_xmin, and lung_xmax, respectively.
  • ctr: cardiothoracic ratio, computed as the ratio of heart width to lung width.

• mediastinal_widening.csv

  • label: 1 indicates the presence of mediastinal widening, 0 indicates its absence.
  • mediastinum_xmin, mediastinum_xmax: X-coordinates of the mediastinum's left and right end points.
  • lung_xmin, lung_xmax: X-coordinates of the lung's left and right end points at the .
  • mediastinum_width, lung_width: Widths of the mediastinum and lungs, calculated from the coordinates mediastinum_xmin, mediastinum_xmax, lung_xmin, and lung_xmax, respectively.
  • mcr: mediastinal to chest width ratio, computed as the ratio of mediastinum width to lung width.

• carina_angle.csv

  • label: 1 indicates an abnormal carina angle, 0 indicates a normal angle.
  • point_1, point_2, point_3: Coordinates of three points defining the tracheal bifurcation, where point_2 is the bifurcation point and point_1 and point_3 are points along the right and left main bronchi, respectively.
  • angle: The angle formed at the bifurcation point, measured directly in degrees.

• trachea_deviation.csv

  • label: 1 indicates trachea deviation, 0 indicates no deviation.
  • point_1 to point_9: Coordinates of nine points along the midline of the trachea.
  • direction_per_pnt: Direction of the trachea mask relative to each point at the same horizontal level (same y-coordinate), labeled as left, right, or flat for each of the nine points.
  • direction: Overall trachea deviation direction, determined by aggregating the per-point direction assessments.

• aortic_knob_enlargement.csv

  • label: 1 indicates enlargement of the aortic knob, 0 indicates no enlargement.
  • aortic_knob_xmin, aortic_knob_xmax: X-coordinates of the aortic knob's left and right end points.
  • trachea_point_left, trachea_point_right: Coordinates of the trachea's left and right end points.
  • aortic_knob_width, trachea_width: Widths of the aortic knob and trachea, calculated from the coordinates aortic_knob_xmin, aortic_knob_xmax, trachea_point_left, and trachea_point_right, respectively.
  • ratio: the ratio of aortic knob width to trachea width.

• ascending_aorta_enlargement.csv

  • label: 1 indicates enlargement of the ascending aorta, 0 indicates no enlargement.
  • heart_point: Coordinates of the most right-lung side point on the right heart border.
  • trachea_point: Coordinates of the most right-lung side point of the trachea.
  • ratio: Ratio of the area of the ascending aorta that extends beyond a straight line connecting the heart point and trachea point to the total area of the ascending aorta mask.

• descending_aorta_enlargement.csv

  • label: 1 indicates enlargement of the descending aorta, 0 indicates no enlargement.
  • desc_aorta_point_left,desc_aorta_point_right: Coordinates of the descending aorta's left and right end points.
  • trachea_point_left, trachea_point_right: Coordinates of the trachea's left and right end points.
  • desc_aorta_width, trachea_width, : Widths of the descending aorta and trachea, calculated from the coordinates desc_aorta_point_left, desc_aorta_point_right, trachea_point_left, and trachea_point_right, respectively.
  • ratio: the ratio of descending aorta width to trachea width.

• descending_aorta_tortuous.csv

  • label: 1 indicates tortuosity of the descending aorta, 0 indicates no tortuosity.
  • point_1 to point_6: Coordinates of the rightmost points at the boundaries of five equal vertical segments of the thoracic descending aorta region.
  • curvature: Mean curvature across the six points, computed using finite difference methods to evaluate the tortuosity.

Image Quality Assessment

• inclusion.csv

  • label: 1 indicates sufficient inclusion of lung regions (e.g., apex, side, and bottom are adequately captured), 0 indicates exclusion.
  • label_side_right_lung, label_side_left_lung, label_apex_right_lung, label_apex_left_lung, label_bottom_right_lung, label_bottom_left_lung: Binary labels (1 for inclusion, 0 for exclusion) indicating whether the side, apex, and bottom of the thoracic cage are included in the image.
  • point_apex_right_lung, point_apex_left_lung, point_side_right_lung, point_side_left_lung, point_bottom_right_lung, point_bottom_left_lung: Coordinates of the apex, side, and bottom of the right and left lungs.
  • ratio_apex_right_lung, ratio_apex_left_lung, ratio_side_right_lung, ratio_side_left_lung, ratio_bottom_right_lung, ratio_bottom_left_lung: Ratios of the distance from each corresponding coordinate (apex, side, bottom) to the image edge, relative to a reference distance, for the right and left lungs.

• inspiration.csv

  • label: 1 indicates inadequate inspiration, 0 indicates adequate inspiration
  • mid-clavicular_line(x): X-coordinate of the mid-clavicular line
  • rib_position: The number of the first right posterior rib that intersects with the right hemidiaphragm at the mid-clavicular line

• rotation.csv

  • label: 1 indicates rotation, 0 indicates proper alignment.
  • medial_end_right_clavicle, medial_end_left_clavicle: Coordinates of the medial end points of the right and left clavicles.
  • midline_points: Coordinates of the spinous process points nearest to each clavicle's medial end point, defined within a ±1 vertebrae margin around the nearest spinous process for both clavicles.
  • midline_point_right_clavicle, midline_point_left_clavicle: Coordinates of the points on the spinous process line at the same vertical level (y-coordinate) as the medial end points of the right and left clavicles, respectively
  • distance_right, distance_left: Horizontal distances between the medial end point of the right and left clavicles and their corresponding points on the spinous process line.
  • ratio: Ratio of the minimum to the maximum of the right and left distances, calculated as min(distance_right, distance_left) / max(distance_right, distance_left).

• projection.csv

  • label: 1 indicates large scapular overlap with lung fields, 0 indicates indicates minimal overlap.
  • scapular_region_right, scapular_region_left: Areas of the right and left scapular regions, derived from segmented masks.
  • overlap_region_righn, overlap_region_left: Areas of overlap between the right and left scapulae and the lung fields.
  • ratio_right, ratio_left: Ratios of the overlap area to the total scapular region area for the right and left scapulae, respectively.

For detailed extraction methods of these columns, refer to Appendix A Details of CheXStruct in the paper CXReasonBench: A Benchmark for Evaluating Structured Diagnostic Reasoning in Chest X-rays.

---

Citation

If you use this dataset in your research, please cite the following paper:

@article{lee2025cxreasonbench,
  title={CXReasonBench: A Benchmark for Evaluating Structured Diagnostic Reasoning in Chest X-rays},
  author={Lee, Hyungyung and Choi, Geon and Lee, Jung-Oh and Yoon, Hangyul and Hong, Hyuk Gi and Choi, Edward},
  journal={arXiv preprint arXiv:2505.18087},
  year={2025}
}

---

License

The dataset is released under CC BY 4.0. Please ensure compliance with the licenses of the source datasets when using this dataset.

---

Contact

For questions or issues related to the dataset, please open an issue on the Hugging Face dataset repository or contact: ttumyche@kaist.ac.kr