Automated Segmentation of Diffuse and Multifocal Nerve Enlargement in Immune-Mediated Neuropathy Using Temporal Deep...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Peripheral nerve ultrasound assesses nerve enlargement in immune-mediated neuropathies, but challenges exist due to variable distribution, severity, indistinct boundaries, and heterogeneous echogenicity.
• Previous deep learning segmentation studies often focused on limited regions or single sites, primarily in compressive neuropathies, not diffuse nerve enlargement in immune-mediated conditions.

Purpose of the Study:

• To evaluate the performance of temporal deep learning-based segmentation for assessing diffuse or focal nerve enlargement in immune-mediated neuropathies.
• To compare static and temporal deep learning models using continuous ultrasound scans for nerve segmentation.

Main Methods:

• Continuous ultrasound scans from wrist to below the elbow were performed on 25 healthy participants and 5 patients with immune-mediated neuropathy.
• Static DeepLabV3+ and temporal models (ConvLSTM, Temporal Mamba) were constructed and compared for nerve segmentation accuracy.

Main Results:

• Temporal models demonstrated higher Dice coefficients and reduced frame-to-frame variability in patients with nerve enlargement compared to healthy participants.
• The ConvLSTM-based temporal model achieved the highest performance, with mean Dice coefficients ranging from 0.87 to 0.92.
• Segmentation performance was comparable across models in healthy participants.

Conclusions:

• Temporal deep learning shows potential for improved nerve segmentation in immune-mediated neuropathies with nerve enlargement.
• Temporal models enhance segmentation consistency and reduce variability, facilitating quantitative ultrasound evaluation.
• Further validation in larger cohorts is warranted for this promising approach.