SCOPE-MRI: Bankart Lesion Detection as a Case Study in Data Curation and Deep Learning for Challenging Diagnoses

Sahil Sethi, Sai Reddy, Mansi Sakarvadia, Jordan Serotte, Darlington Nwaudo, Nicholas Maassen, Lewis ShiarXiv 2025 · DOI · arXiv

Abstract

This paper introduces ScopeMRI, the first publicly available, expert-annotated dataset for shoulder pathologies, and presents a deep learning framework specifically designed for detecting subtle Bankart lesions. By leveraging domain-specific pretraining and multi-view ensembling, the proposed models achieve radiologist-level performance on non-invasive standard MRIs, potentially matching the diagnostic accuracy of invasive, contrast-enhanced procedures.

Introduction

In musculoskeletal (MSK) imaging, deep learning (DL) has successfully automated many "easy" tasks, such as detecting ACL tears. However, Bankart lesions—tears in the shoulder's anterior-inferior labrum—remain a significant challenge. These lesions are a primary cause of chronic shoulder instability but are notoriously difficult to identify on standard, non-contrast MRIs due to their subtle imaging features. Consequently, radiologists often achieve only ~50% sensitivity on standard scans, frequently necessitating an MRI Arthrogram (MRA). MRAs are invasive, expensive, and uncomfortable, requiring direct joint injection of contrast dye. This research aims to bridge the gap: can AI detect these subtle tears on standard MRIs effectively enough to reduce the need for invasive procedures?

Methods

The researchers developed a comprehensive workflow optimized for small and imbalanced medical datasets:

  1. Architecture Selection: The team evaluated various architectures, including traditional Convolutional Neural Networks (AlexNet) and modern Transformers (Vision Transformer and Swin Transformer).
  2. Domain-Relevant Transfer Learning: Instead of generic ImageNet pretraining, models were pretrained on MRNet, a large dataset of knee MRIs. This "medical-to-medical" transfer significantly boosted performance.
  3. Multi-View Ensembling: To mimic clinical diagnosis, separate models were trained for Sagittal, Axial, and Coronal views and then ensembled.
  4. Stability-Based Selection: 8-fold stratified cross-validation was used to identify the most stable architecture with the lowest variance across data splits.
  5. Interpretability: Grad-CAM heatmaps were utilized to ensure the model focused on the labrum rather than unrelated image artifacts.

Datasets

  • ScopeMRI (Primary): A newly released dataset of 586 shoulder MRIs (335 standard, 251 MRA) from 546 patients. Ground-truth labels were derived from intraoperative findings (arthroscopy), the gold standard for diagnosis.
  • External Dataset: 12 standard MRIs from independent hospitals were used to test model generalization across different hardware and protocols.
  • MRNet: Utilized for domain-specific pretraining.

Key Insights: The Power of Medical Pretraining

A critical finding was the substantial performance gain from MRNet pretraining compared to generic ImageNet pretraining. For the challenging standard MRI task, medical-specific pretraining improved the Area Under the Curve (AUC) by nearly 9%.

Heatmap showing the AUC improvement (%) gained by using MRNet pretraining over ImageNet pretraining across different models and views.

Furthermore, the study revealed that model complexity should match task difficulty. Simpler models like AlexNet performed well on high-contrast MRA views, while more complex Transformers (Swin) were necessary to extract features from the more difficult Sagittal and Coronal standard MRI views.

Key Results

The ensemble model demonstrated remarkable performance, particularly on non-invasive scans:

  • Standard MRI: The ensemble achieved an AUC of 0.91, surpassing the sensitivity of original radiology reports (83.33% vs. 16.67%).
  • MRA Performance: On higher-contrast MRA scans, the model reached an AUC of 0.93.
  • Generalizability: The model maintained a 0.90 AUC on the external dataset, indicating robust feature learning across different clinical environments.
ROC curves for internal standard MRI, internal MRA, and external standard MRI datasets, comparing ensemble performance to individual views and radiologist benchmarks.

Limitations and Future Work

While promising, the study acknowledges several hurdles, including a small positive sample size in the standard MRI test set and significant dataset imbalance (only 8.6% prevalence of Bankart lesions). Future work will explore "virtual staining"—using AI to synthesize MRA-like clarity from standard MRIs—and conducting reader studies to evaluate how AI insights impact real-time radiologist accuracy.

Code and Data Availability: The authors have released the ScopeMRI dataset and their modular codebase on GitHub to support ongoing research.

Summary updated

BibTeX
@misc{sethi2025scope,
  author = {Sahil Sethi and Sai Reddy and Mansi Sakarvadia and Jordan Serotte and Darlington Nwaudo and Nicholas Maassen and Lewis Shi},
  title = {{SCOPE-MRI: Bankart Lesion Detection as a Case Study in Data Curation and Deep Learning for Challenging Diagnoses}},
  year = {2025},
  eprint = {2504.20405},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.20405},
}
Figure