Brain Hemorrhage Extended (BHX): Bounding Box Extrapolation from Thick to Thin Slice CT Images

Creators:
Eduardo Pontes Reis, Felipe Nascimento, Mateus Aranha, Fernando Mainetti Secol, Birajara Machado, Marcelo Felix, Anouk Stein, Edson Amaro

Published: July 29, 2020
Version: 1.1

📄 Citation

Reis, E. P., Nascimento, F., Aranha, M., Mainetti Secol, F., Machado, B., Felix, M., Stein, A., & Amaro, E. (2020). Brain Hemorrhage Extended (BHX): Bounding box extrapolation from thick to thin slice CT images (version 1.1). PhysioNet. https://doi.org/10.13026/9cft-hg92

PhysioNet Standard Citation:

Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov, P. C., Mark, R., ... & Stanley, H. E. (2000). PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation [Online]. 101 (23), pp. e215–e220.

🧠 Abstract

BHX is a publicly available dataset with bounding box annotations for five types of acute hemorrhage as an extension of the qure.ai CQ500 dataset. The dataset is intended to aid in developing machine learning solutions for hemorrhage detection and localization.

📌 Key Points

39,668 bounding boxes in 23,409 images annotated for hemorrhage.
Built from the ~170k image CQ500 dataset.
Bounding boxes extrapolated from sparse thick-slice labeling to thin-slice images.
Supports machine learning applications for hemorrhage localization and diagnosis.

🧬 Background

Intracranial hemorrhage is a serious condition with a 40% one-month mortality rate.
Head CT is the primary imaging modality.
Manual bounding box annotation is time-consuming; hence, extrapolation methods are used.
No existing public dataset includes localization data with bounding boxes until BHX.

🧪 Methods

Based on the CQ500 dataset (491 scans, 205 hemorrhage-positive).
Labeled by 3 neuroradiologists with varying experience.
Matching between thick and thin series was done via DICOM tag: Image Position (patient).
Six types of hemorrhage labeled:
Intraparenchymal
Subarachnoid
Intraventricular
Epidural
Acute Subdural
Chronic Subdural

🗃️ Data Description

Annotation Stats:

6,283 manually labeled bounding boxes in 3,558 images.
39,668 extrapolated bounding boxes in 23,409 images.

Dataset Versions:

1_Initial_Manual_Labeling.csv: Hand-drawn annotations on thick slices.
2_Extrapolation_to_All_Series.csv: Extrapolated to all corresponding series.
3_Extrapolation_to_Selected_Series.csv: Extrapolated only for selected soft-tissue thin-slice series.

Columns:

SOPInstanceUID: Unique DICOM image ID
SeriesInstanceUID: DICOM series ID
StudyInstanceUID: DICOM study ID
data: Bounding box coordinates (X, Y, width, height)
labelName: Hemorrhage type
labelType: Source of image (thick-slices, thin-slices, or other)

🔗 DICOM UID Mapping

Annotations are linked via DICOM tag: 0008,0018 – SOP Instance UID.

📂 Original Images

Hosted at: http://headctstudy.qure.ai/dataset

🔍 Usage Notes

Unique resource for bounding-box annotated hemorrhage images.
Enables benchmarking and development of deep learning algorithms.
Includes extrapolated labels, some of which may have minor inaccuracies.
Future work should consider interpolating bounding boxes between slices for smoother transitions.

🖼️ Visual Inspection

Explore annotated images at:
https://public.md.ai/annotator/project/Y2qr6vqv/workspace

🙏 Acknowledgements

qure.ai – for publishing the CQ500 dataset.
MD.ai – for providing the annotation platform.

⚠️ Conflicts of Interest

A.S. is employed by MD.ai, which provided the annotation platform.

📚 References

van Asch C, et al. The Lancet Neurology, 2010.
Heit J, et al. Journal of Stroke, 2017.
Chang P, et al. American Journal of Neuroradiology, 2018.
Goldstein J, Gilson A. Current Treatment Options in Neurology, 2011.
Prevedello L, et al. Radiology, 2017.
Chilamkurthy S, et al. The Lancet, 2018.
Kuo W, et al. PNAS, 2019.
RSNA ICH Detection. Kaggle
Mirza S, Gokhale S. IntechOpen, 2017.
Osborn A, et al. Osborn’s Brain, Elsevier, 2018.
Weiss K, et al. AJR, 2011.
DICOM Standard. dicomstandard.org