A deep learning framework for automated quantification of peripheral nerve lesions in MR neurography

Abstract

Magnetic resonance neurography (MRN) enables high-resolution visualization of peripheral nerves, yet quantitative analysis of intraneural lesion burden remains limited by manual, observer-dependent workflows. We propose a deep learning-based framework for automated segmentation and quantitative characterization of peripheral nerve lesions in MRN. In this retrospective study (n = 178), standardized axial fat-suppressed T2-weighted 3-T MRN scans of the distal thigh were analyzed. Sciatic nerves and intraneural lesions were manually annotated to train a modified nnU-Net v2 architecture for 2D and 3D segmentation. Model performance was evaluated using Dice and surface Dice metrics. Generalizability was assessed across two independent test datasets. Imaging-derived lesion burden metrics were further related to electrophysiological parameters using age-adjusted partial correlation analyses. Automated nerve segmentation achieved a median Dice coefficient of 0.97. Lesion segmentation yielded surface Dice values of 0.64–0.77 across training and independent test sets, demonstrating stable generalization in a small-structure segmentation setting. Quantitative lesion burden metrics were consistent across cohorts and showed significant associations with electrophysiological measures of nerve function. For example, automated total lesion load correlated significantly with tibial nerve conduction velocity (r = -0.62, p < 0.001) and tibial compound motor action potential (r = -0.568, p 0.002) in age-adjusted partial correlation analyses. This work introduces a reproducible deep learning-based framework for automated segmentation and quantitative assessment of peripheral nerve lesions in MR neurography. The approach enables standardized volumetric characterization of small intraneural abnormalities and provides a basis for scalable, quantitative peripheral nerve imaging. Question: AI automatically detects and measures nerve damage in MRI scans. Findings: Method provides a fast, consistent alternative to manual image analysis. Relevance statement: Automated MRI-based nerve lesion quantification enables objective, reproducible assessment of peripheral neuropathies, potentially improving diagnostic accuracy, monitoring disease progression, and supporting standardized, scalable evaluation in both clinical practice and research. Question: AI automatically detects and measures nerve damage in MRI scans. Findings: Method provides a fast, consistent alternative to manual image analysis. Relevance statement: Automated MRI-based nerve lesion quantification enables objective, reproducible assessment of peripheral neuropathies, potentially improving diagnostic accuracy, monitoring disease progression, and supporting standardized, scalable evaluation in both clinical practice and research.