Deep Learning Accurately Quantifies Peripheral Nerve Lesions in MR Neurography

In a retrospective computational neuroimaging investigation encompassing one hundred and seventy-eight subjects, researchers have validated a deep learning pipeline capable of automatically delineating and volumetrically quantifying small intraneural lesions within the sciatic nerve on magnetic resonance neurography scans. The artificial intelligence framework, built upon a systematically adapted nnU-Net v2 architecture, demonstrated robust segmentation fidelity across both healthy individuals and patients diagnosed with distinct peripheral neuropathies. Across training and independent external testing cohorts, the automated nerve segmentation achieved an outstanding median Dice similarity coefficient of 0.97, while the more challenging task of segmenting focal intraneural lesions yielded surface Dice scores ranging between 0.64 and 0.77. Beyond establishing geometric accuracy on high-resolution imaging, the derived quantitative lesion burden metrics demonstrated strong, statistically significant correlations with objective clinical electrophysiological measurements of nerve dysfunction, highlighting the potential of automated volumetric magnetic resonance neurography to serve as a reproducible, standardized structural biomarker for peripheral nerve pathology.

Magnetic resonance neurography has emerged as an indispensable specialized diagnostic modality, utilizing fat-suppressed T2-weighted sequences to visualize peripheral nerve trunks and lumbosacral plexuses with exceptional anatomical resolution. By suppressing background signal from adjacent adipose tissue, these sequences highlight morphological signal alterations, distinguishing between healthy nerve parenchyma and pathological tissue. Clinically, magnetic resonance neurography proves vital in diagnosing inflammatory, metabolic, hereditary, and post-traumatic neuropathies, as well as peripheral nerve sheath tumors, particularly in proximal nerve regions such as the sciatic nerve where traditional surface electrophysiological nerve conduction studies face substantial anatomical access limitations. In diffuse polyneuropathies, focal hyperintense and hypointense signal abnormalities within the sciatic nerve reflect underlying axonal degeneration, demyelination, and structural remodeling. However, translating these qualitative visual impressions into objective, standardized clinical biomarkers has historically been impeded by the labor-intensive nature of manual slice-by-slice contouring, pervasive inter-observer variability, and the difficulty of distinguishing genuine microscopic intraneural lesions from partial-volume effects, adjacent vessels, and regional coil sensitivity variations.

To overcome these longstanding translational barriers, the research team assembled a comprehensive multicenter imaging dataset comprising 178 subjects, partitioned into three methodologically distinct cohorts approved by the institutional ethics committee of Heidelberg University. The model training dataset consisted of 91 participants, including healthy control subjects and patients diagnosed with diabetic neuropathy secondary to type 2 diabetes mellitus displaying characteristic intraneural neuropathic lesions. To rigorously assess the generalizability of the trained convolutional networks across disparate clinical populations and etiologies, the investigators evaluated two independent test datasets. Test dataset 1 gathered 56 subjects presenting a broad spectrum of neuromuscular conditions, encompassing diabetic neuropathy, healthy controls, peripheral nerve schwannomas associated with schwannomatosis and neurofibromatosis type 2, Fabry disease, and amyotrophic lateral sclerosis. Test dataset 2 comprised 31 individuals, consisting of patients with type 1 diabetes-associated diabetic neuropathy, patients presenting with uremic neuropathy, and healthy control participants, all of whom had undergone detailed standardized motor and sensory nerve conduction studies to enable formal structure-function correlation analyses.

All diagnostic imaging examinations were executed on clinical 3.0-Tesla whole-body magnetic resonance scanners manufactured by Siemens Healthineers, utilizing dedicated fifteen-channel transmit-and-receive knee array coils to maintain high signal-to-noise ratios. Image acquisitions were performed across specialized diagnostic environments at the Department of Neuroradiology within the University Hospital Heidelberg, which operated Magnetom Trio and Magnetom Prisma systems, and at the Neuer Wall magnetic resonance imaging facility, which utilized a Magnetom Skyra platform. The acquisition protocol adhered strictly to standardized axial fat-suppressed T2-weighted turbo spin-echo sequences targeted specifically to the distal thigh region. By enforcing standardized technical acquisition parameters across multiple scanner models and institutional environments, the investigators established an imaging baseline designed to reflect the real-world operational demands of advanced neuroradiology centers, while ensuring that the deep learning models learned intrinsic morphological features rather than vendor-specific or coil-specific image acquisition artifacts.

Establishing reliable ground-truth annotations for network training and comparative evaluation required an exhaustive, semi-automated expert labeling procedure executed independently by two experienced readers, consisting of a board-certified radiologist with more than five years of specialized neurography experience and an advanced radiology resident with over three years of practice. Using ImageJ software version 1.53k, the readers initially delineated the sciatic nerve boundaries on each sequential axial slice, restricting subsequent computational analysis strictly to the intraneural compartment. Within this delineated contour, readers established slice-specific grayscale thresholds on converted eight-bit images to extract binary masks of T2-weighted hyperintense and hypointense lesions, avoiding rigid global intensity cutoffs that fail to account for inter-individual tissue contrast and spatial coil sensitivity profiles. Surrounding subcutaneous fat served as the internal anatomical baseline for hypointense lesions, whereas adjacent normal muscle tissue and unaffected nerve fascicles provided the reference standard for hyperintense lesions. Readers systematically reviewed and manually edited all generated masks to eliminate spurious single-slice noise and vascular flow phenomena, ultimately achieving an outstanding interrater intraclass correlation coefficient of 0.93.

The algorithmic architecture implemented by the researchers leveraged a heavily modified nnU-Net v2 framework, training separate two-dimensional and three-dimensional full-resolution convolutional models across one thousand training epochs. A dedicated nerve segmentation model operated across the complete magnetic resonance field of view, utilizing an eight-stage encoder for two-dimensional inputs and a seven-stage encoder for three-dimensional volumes. To address the extreme class imbalance and microscopic scale of intraneural signal alterations, the specialized lesion segmentation model operated on tightly cropped nerve regions of interest, employing four encoder stages. Crucially, the training loss function combined focal loss and soft Dice loss, assigning aggressive loss weights of 2.0 to hyperintense and hypointense lesion voxels versus 0.05 to background tissue. The investigators additionally integrated a top-k voxel mining strategy that concentrated gradient updates exclusively on the ten percent highest-error voxels within each training batch, while test-time augmentation across eight mirroring axes and post-processing filters eliminating isolated lesion clusters smaller than ten cubic millimeters ensured geometric stability.

Evaluation across the external cohorts confirmed the remarkable technical resilience and anatomical precision of the deep learning pipeline. The automated nerve delineation model achieved an exceptional median Dice coefficient of 0.97 and a median surface Dice score of 1.00 in the test datasets, effectively matching human expert contours. For the intraneural lesion segmentation task, median surface Dice coefficients ranged between 0.687 and 0.771 across distinct lesion categories, while conventional volumetric Dice scores spanned from 0.613 to 0.731. When the researchers tested the complete automated pipeline end-to-end, feeding automatically generated nerve masks into the downstream lesion model rather than expert manual nerve delineations, segmentation performance showed only minimal attenuation. Slight local discrepancies were primarily confined to the tibial division of the sciatic nerve in select cases where adjacent vascular structures introduced local boundary ambiguity, or where an absence of intervening fat tissue between the nerve epineurium and adjacent skeletal muscle hindered automated edge distinction.

Volumetric lesion profiling across both independent validation cohorts revealed highly uniform structural patterns of neuropathic tissue involvement. In test dataset 1, the median hyperintense lesion volume stood at 73.7 cubic millimeters per subject, while hypointense lesions accounted for a median volume of 89.1 cubic millimeters, yielding a combined median total lesion volume of 179.9 cubic millimeters. When normalized against total sciatic nerve volume, the median hyperintense lesion fraction was 0.063 and the hypointense fraction was 0.076, producing an aggregate lesion-to-nerve volume ratio of 0.143. In test dataset 2, where neuropathic disease severity was marginally higher, median hyperintense and hypointense lesion volumes measured 94.7 cubic millimeters and 121.9 cubic millimeters, respectively, resulting in a median total lesion burden of 223.3 cubic millimeters. Normalizing for nerve caliber yielded a median hyperintense fraction of 0.067, a hypointense fraction of 0.077, and an overall lesion-to-nerve ratio of 0.154, demonstrating that relative lesion proportions remained remarkably consistent across distinct patient populations.

Detailed diagnostic performance metrics revealed a highly conservative lesion detection profile, characterized by exceptionally high positive predictive values coupled with moderate sensitivity. In test dataset 1 using the fully automated pipeline, the model achieved a median positive predictive value of 0.892 for hyperintense lesions and 0.805 for hypointense lesions, alongside sensitivities of 0.489 and 0.394, respectively, confirming that the network rarely produces false-positive detections despite omitting a proportion of subtle lesion boundaries. In the electrophysiological validation conducted on test dataset 2, age-adjusted partial correlation analyses demonstrated that automatically quantified total lesion burden correlated strongly and inversely with tibial nerve conduction velocity, exhibiting a correlation coefficient of negative 0.62 with a p-value below 0.001. Furthermore, total lesion volume displayed a statistically significant negative association with tibial compound muscle action potential amplitudes, yielding a correlation coefficient of negative 0.568 with a p-value of 0.002, demonstrating that automated volumetric measurements parallel and sometimes exceed manual quantification in capturing clinically meaningful neurophysiological deterioration.

Despite these encouraging findings, the authors emphasized several methodological and translational limitations that must be addressed prior to clinical implementation. The investigation was retrospective in nature, introducing potential selection bias, and relied on a training cohort predominantly composed of patients with type 2 diabetic neuropathy. While external testing across diverse non-diabetic conditions demonstrated substantial disease-independent robustness, larger prospective cohorts remain necessary. Furthermore, all imaging data were acquired exclusively on 3.0-Tesla scanners from a single manufacturer using standardized fat-suppression parameters; network generalizability across lower field strengths, alternative vendor platforms, or different fat-suppression techniques such as short tau inversion recovery remains uncharacterized. The authors also noted that eight-bit image conversion during thresholding may have compressed subtle contrast gradations, that tiny intraneural vessels could occasionally confound hyperintense lesion boundaries, and that the modest sample size of test dataset 2 without multiple comparison adjustments requires treating neurophysiological correlations as exploratory rather than confirmatory. Consequently, the framework serves as a validated technical foundation for future multicenter trials rather than an immediate diagnostic tool.

Subject of Research: Deep learning-based automated segmentation and volumetric quantification of peripheral nerve lesions in magnetic resonance neurography

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

Article References: Beste, N. C., Raudonat, C., Fesselier, M., Hartmann, S., Sommer, D., Jende, J. M. E., Kottlors, J., Heiland, S., Meredig, H., Bendszus, M., & Mooshage, C. M. (2026). A deep learning framework for automated quantification of peripheral nerve lesions in MR neurography. European Radiology Experimental, 10(1), Article 122. https://doi.org/10.1186/s41747-026-00786-z

Image Credits: AI Generated

DOI: 10.1186/s41747-026-00786-z

Keywords: Magnetic Resonance Neurography, Peripheral Neuropathy, Sciatic Nerve, nnU-Net, Deep Learning, Convolutional Neural Networks, Image Segmentation, Diabetic Neuropathy, Electrophysiology, Nerve Conduction Velocity, Compound Muscle Action Potential, Surface Dice Metric, Intraneural Lesions, T2-Weighted MRI, Quantitative Biomarkers, Interrater Reliability, Schwannomatosis, Fabry Disease