Non-traumatic brachial plexopathy identification from routine MRIs: Retrospective studies with deep learning networks

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Brachial plexopathy (BP) diagnosis can be challenging using standard MRI.
• Deep learning (DL) offers potential for improving diagnostic accuracy in medical imaging.

Purpose of the Study:

• To optimize a deep learning model for differentiating non-traumatic brachial plexopathy from routine MRI scans.
• To evaluate the performance of various DL schemes across multiple MRI sequences.

Main Methods:

• Retrospective analysis of 256 brachial plexus MRI series (196 patients) from Mayo Clinic (2002-2022).
• Six DL schemes were designed using six DL networks as backbones, trained and validated via nested five-fold cross-validation.
• Three MRI sequences (sagittal T1, fluid-sensitive, post-gadolinium) were utilized, with regions of interest (ROIs) selected by a radiology panel.

Main Results:

• The best-performing model, utilizing feature merging with a triple MRI joint strategy, achieved an AUC of 92.2% and accuracy of 89.5%.
• This outperformed other models, including multiple channel merging (AUC 89.6%) and solo channel volume (AUC 89.2%).
• The feature merging mode with VGG16 backbone achieved the highest method score (1.75/2.37).

Conclusions:

• Deep learning models show significant potential for diagnosing brachial plexopathy using combined MRI sequences.
• Feature merging and multi-sequence analysis provide a superior approach for DL-based BP abnormality detection compared to single-sequence analysis.