sahilsethi0105/scope-mri

ScopeMRI

Dataset Summary

ScopeMRI (Shoulder Comprehensive Orthopedic Pathology Evaluation–MRI) is a medical imaging dataset designed to support machine learning research in musculoskeletal MRI interpretation.

The dataset contains 586 shoulder MRI studies from 558 patients at the University of Chicago, including:

• 335 standard shoulder MRI scans
• 251 MRI arthrograms (MRA) (MRI scans with intra-articular contrast)

The dataset was developed to study automated detection of shoulder labral and rotator cuff pathology, with labels derived from arthroscopic surgical findings, which represent the clinical gold standard for diagnosis.

ScopeMRI was introduced in the paper:

SCOPE-MRI: Bankart Lesion Detection as a Case Study in Data Curation and Deep Learning for Challenging Diagnoses
https://www.nature.com/articles/s44387-025-00043-5

---

Dataset Access

The imaging data are hosted in the Medical Imaging and Data Resource Center (MIDRC).

• MIDRC Data Commons: https://www.midrc.org/
• Repository with dataset documentation and access instructions:
  https://github.com/sahilsethi0105/scope-mri/

The imaging files themselves are not stored in this Hugging Face repository.
Users must follow the instructions in the repository to request and download the dataset from MIDRC.

---

Dataset Composition

Imaging Data

• Total shoulders: 586
• MRI arthrograms (MRA): 251
• Standard MRI: 335
• Image format: DICOM

Each study consists of several standad MRI sequences acquired in clinical practice (including T1/T2 in different views including sagittal, axial, etc.).

---

Metadata

The dataset includes associated metadata for each study:

• Age
• Sex
• Race
• Ethnicity
• 3-digit ZIP code
• Imaging acquisition parameters

Acquisition parameters are available in both:

• DICOM headers
• Metadata spreadsheets

---

Labels

Binary labels are provided for the following shoulder pathologies:

• Bankart lesion
• Reverse Bankart lesion
• Superior labrum tear
• Rotator cuff tear (any type)

Labels were determined using arthroscopic surgical findings, which serve as the clinical ground truth.

---

Data Format

• Imaging files are stored in DICOM format.
• Additional metadata and labels are provided in tabular spreadsheets.

---

Data Splits

No predefined training, validation, or test splits are provided with the dataset.

Researchers are expected to create their own splits depending on their experimental design.

---

Data Anonymization

All imaging data were de-identified prior to release.

---

Intended Use

ScopeMRI is intended for research in:

• Medical image analysis
• Machine learning for MRI interpretation
• Automated detection of shoulder pathology
• Clinical decision support systems for musculoskeletal imaging

---

Limitations

• The dataset represents patients who underwent arthroscopic surgery, which may introduce selection bias.
• All scans originate from a single institution.
• No official dataset splits are provided.

---

Citation

If you use ScopeMRI in your research, please cite:

@article{sethi_scope-mri_2025,
    title = {{SCOPE}-{MRI}: {Bankart} lesion detection as a case study in data curation and deep learning for challenging diagnoses},
    volume = {1},
    issn = {3005-1460},
    url = {https://doi.org/10.1038/s44387-025-00043-5},
    doi = {10.1038/s44387-025-00043-5},
    abstract = {Deep learning has shown strong performance in musculoskeletal imaging, but prior work has largely targeted conditions where diagnosis is relatively straightforward. More challenging problems remain underexplored, such as detecting Bankart lesions (anterior-inferior glenoid labral tears) on standard MRIs. These lesions are difficult to diagnose due to subtle imaging features, often necessitating invasive MRI arthrograms (MRAs). We introduce ScopeMRI, the first publicly available, expert-annotated dataset for shoulder pathologies, and present a deep learning framework for Bankart lesion detection on both standard MRIs and MRAs. ScopeMRI contains shoulder MRIs from patients who underwent arthroscopy, providing ground-truth labels from intraoperative findings, the diagnostic gold standard. Separate models were trained for MRIs and MRAs using CNN- and transformer-based architectures, with predictions ensembled across multiple imaging planes. Our models achieved radiologist-level performance, with accuracy on standard MRIs surpassing radiologists interpreting MRAs. External validation on independent hospital data demonstrated initial generalizability across imaging protocols. By releasing ScopeMRI and a modular codebase for training and evaluation, we aim to accelerate research in musculoskeletal imaging and foster the development of datasets and models that address clinically challenging diagnostic tasks.},
    number = {1},
    journal = {npj Artificial Intelligence},
    author = {Sethi, Sahil and Reddy, Sai and Sakarvadia, Mansi and Serotte, Jordan and Nwaudo, Darlington and Maassen, Nicholas and Shi, Lewis},
    month = dec,
    year = {2025},
    pages = {41},
}

and

@inproceedings{sethi_toward_2025,
    title = {Toward non-invasive diagnosis of {Bankart} lesions with deep learning},
    volume = {13407},
    url = {https://www.spiedigitallibrary.org/conference-proceedings-of-spie/13407/134073G/Toward-non-invasive-diagnosis-of-Bankart-lesions-with-deep-learning/10.1117/12.3046251.full},
    doi = {10.1117/12.3046251},
    booktitle = {Medical {Imaging} 2025: {Computer}-{Aided} {Diagnosis}},
    publisher = {SPIE},
    author = {Sethi, Sahil and Reddy, Sai and Sakarvadia, Mansi and Serotte, Jordan and Nwaudo, Darlington and Maassen, Nicholas and Shi, Lewis},
    month = apr,
    year = {2025},
    pages = {854--863},
}