Special Poster Session 51st International Society for the Study of the Lumbar Spine Annual Meeting 2025

Automatic generation of diffusion tensor imaging for the lumbar nerve using convolutional neural networks (115360)

Rira Masumoto 1 , Yawara Eguchi 2 3 , Hidenari Takeuchi 4 , Kazuhide Inage 2 , Miyako Narita 2 , Yasuhiro Shiga 2 , Masahiro Inoue 2 , Noriyasu Toshi 2 , Soichiro Tokeshi 2 , Kohei Okuyama 2 , Shuhei Ohyama 2 , Noritaka Suzuki 1 , Satoshi Maki 2 5 , Takeo Furuya 2 , Seiji Ohtori 2 , Sumihisa Orita 1 2 5
  1. Department of Medical Engineering, Graduate School of Science and Engineering, Chiba University, Japan, Chiba, Japan
  2. Department of Orthopaedic Surgery, Graduate School of Medicine,, Chiba University, Japan, Chiba, Japan
  3. Department of Orthopaedic Surgery, Shimoshizu National Hospital, Yotsukaido, Chiba, Japan
  4. Works Human Intelligence, Tokyo
  5. Center for Frontier Medical Engineering, Chiba University, Japan, Chiba, Japan

6737029cc3dce-%E3%82%B9%E3%82%AF%E3%83%AA%E3%83%BC%E3%83%B3%E3%82%B7%E3%83%A7%E3%83%83%E3%83%88+2024-11-15+17.10.39.png

Purpose
One of the primary causes of low back pain and leg pain in lumbar spine disorders is lumbar foraminal stenosis, resulting from compression of the lumbar nerve roots. While MRI (Magnetic Resonance Imaging) is widely used for morphological evaluation of such nerve impairments, it remains challenging to visualize the functional pathologies of the nerve roots. Lumbar nerve root tractography, developed using Diffusion Tensor Imaging (DTI), focuses on the diffusion of water molecules and utilizes Fractional Anisotropy (FA), a measure of diffusion anisotropy, to visualize the nerve impairment causing pain. However, this technique involves manual visual assessments, posing challenges in ensuring reproducibility. Previous studies have introduced deep learning-based automatic segmentation methods, enabling automatic tractography generation without location constraints. Nevertheless, the accuracy of the generated tractographies was insufficient for clinical applications. This study aimed to improve the segmentation accuracy of nerve root regions and develop a system for generating clearer and more reliable tractographies automatically.

Methods
DTI data from 90 patients with lumbar degenerative diseases were used to extract 839 diffusion-weighted images (DWIs). Annotations were made at three regions: two corresponding to the L3-L5 lumbar nerve root locations and one to the dural canal. Separate labels for nerve roots and the dural canal were used during annotation to improve segmentation accuracy. These annotated images were used as training data for U-Net, a semantic segmentation model, to predict and output mask images of nerve root regions for new DTI inputs. The segmentation accuracy of five different architectural models was compared using the Dice coefficient. Additionally, the tractographies generated by the model were evaluated to determine whether they could more accurately trace nerves compared to those generated in previous studies.

Results
Among the five models, the Resnet34 architecture achieved the best results, with a Dice coefficient of 0.780. However, a previous similar study that performed segmentation of the spinal cord region using MRI reported a Dice coefficient of 0.91, indicating that the accuracy achieved in this study was comparatively lower. Furthermore, the tractographies generated by the proposed model were qualitatively evaluated and found to more accurately trace the nerve pathways compared to those from previous studies.

Discussion
The current study successfully developed a system for automatically extracting lumbar nerves and generating tractographies using deep learning-based semantic segmentation. However, the terminal portions of the nerves were often not recognized in the images, highlighting limitations in detecting fine neural structures. These facts indicate that further improvements in accuracy are essential for clinical applications. Future research should focus on analyzing vector information from diffusion anisotropy obtained through DTI to capture the detailed characteristics of nerve pathways more accurately. This approach is surely expected to improve auto-generated tractography accuracy and enable more reliable assessments of neural impairments.