Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Neural Engineering

Neural Engineering - Poster Session C

Poster K3 - Image Processing of X-rays of the Lumbar Spine and Spinal Cord Stimulation Implants

Friday, October 25, 2024
10:00 AM - 11:00 AM EST
Location: Exhibit Hall E, F & G

Introduction:  Neural injuries, such as spinal cord injury (SCI) and stroke, lead to significant motor impairments affecting individuals' quality of life. Motor neuro-prostheses can vastly improve rehabilitation outcomes and increase independence and mobility. Spinal cord stimulation (SCS) has been FDA-approved for chronic pain management [1] and was recently investigated in clinical trials for SCI rehabilitation, particularly for enabling locomotion [2-4]. When SCS is used to treat pain, electrodes are typically placed over the dorsal columns, while for motor function recovery, electrodes are placed more laterally, targeting the dorsal roots.

The location of the electrodes is determined intraoperatively using fluoroscopy, with confirmation of electrode location relative to the vertebral segments in X-ray images. Of interest is pairing the anatomical electrode location with the functional output, measured as muscle activity using electromyography (EMG). However, current methods of manually identifying lead placement relative to the spinal vertebral level via X-ray images are time-consuming and prone to inaccuracies. This compromises the ability to accurately correlate muscle activity from EMG with the location of the corresponding SCS electrode. To overcome this challenge, we propose to use image analysis and machine learning methods to automate the detection of electrode location relative to the spinal vertebral level. Additionally, we will test a new electrode approach in cadavers to better target motor function. Automated image processing will enable us to precisely correlate evoked motor responses with specific spinal cord levels where the stimulation was applied, thus enhancing the effectiveness of SCS for targeted muscle activation and improving rehabilitation outcomes.

Materials and

Methods:  We searched through medical records to construct a dataset of 200 X-ray images from patients who received SCS implant in the lumbosacral region. The dataset was divided into training (80%) and testing (20%) groups. All images were preprocessed, including resizing, cropping, and contrast enhancement as needed. For training annotations, pixel labels were used to label T11-12 and L1-5 vertebrae, and the individual electrodes of the array. This process involved manually outlining each vertebra and electrode using MATLAB’s Image Labeling App, which generated a ground truth dataset for training the network. We constructed and trained a U-Net in MATLAB. A U-Net is a type of convolutional neural network that automatically segments images. We used a U-Net to segment the vertebrae and electrodes from the X-ray images. The trained CNN’s performance was evaluated on the testing subset to ensure generalizability, with fine-tuning as needed to improve performance.

In one human cadaver, we tested the implant method of a new SCS electrode that we hypothesize will achieve improved motor function compared to standard clinical leads developed for pain. Our novel SCS leads were implanted into the epidural space of the spinal cord under fluoroscopic guidance by trained clinicians, then steered laterally to better target the roots. Through cadaver testing, we demonstrated feasibility and refined the implant procedure. Safety was determined by examining the tissue during careful dissection by an anatomist. Following the implant procedure in the cadaver, we used our trained U-Net to automatically segment the X-ray images obtained from fluoroscopy.

Results, Conclusions, and Discussions: We annotated sample X-ray images of the lumbar spinal cord with implanted SCS electrodes and exported them for training purposes. Images were pre-processed to ensure that they were of the same size for training the U-Net. A preliminary cadaver dissection was performed by an anatomist to visualize the spinal canal using an anterior approach and spondylectomy. The anatomist exposed the spinal canal by removing the psoas muscle attachment to the L1 vertebra and the vertebral body, allowing us to visualize it. This initial dissection provided a guide for implant testing in a human cadaver.

Following successful implantation, anterior-posterior X-ray images of the implanted lead were obtained and input into the trained U-Net for automatic identification of electrodes and vertebrae. Three additional cadaver implants are planned to obtain more images for analysis by the trained U-Net.

Following the cadaver testing, the lead placement will be tested intraoperatively, and motor evoked responses will be recorded with surface EMG. The location of the electrodes will be imaged during fluoroscopic guidance to ensure accurate placement.

We will use our trained U-Net to automatically detect and annotate individual electrodes and vertebrae from images obtained during intraoperative testing. This intraoperative testing, combined with our automated identification of electrodes relative to vertebra, will be used to correlate the anatomical position of the electrodes with the muscle activity from EMG. Recording EMG during these intraoperative procedures will allow us to construct a functional map of muscle activation in response to SCS. As SCS becomes more widely used as a rehabilitation intervention, the need for creating a functional map between electrode location and muscle responses will increase. This work will be a valuable tool to researchers and clinicians in automatically generating a functional map from SCS implants.

References: [1] J. Caylor et al., 2019. [2] A. Rowald et al., 2022. [3] F. B. Wagner et al., 2018. [4] M. L. Gill et al., 2018.

Acknowledgements (Optional):