Automated Spinal Cord Segmentation Pipeline Based on Convolutional Neural Network (CNN) for Individualized Cat Spinal Cord Stimulation Modeling

Background and Research Motivation

Spinal cord stimulation (SCS) is a widely used treatment method for chronic pain management. In recent years, it has also been used to modulate neural activity, aiming to restore lost autonomic or sensory-motor functions. Personalized modeling and treatment planning are important aspects to ensure the safety and effectiveness of SCS. However, generating spinal models with the required level of detail and accuracy requires time-consuming and labor-intensive manual image segmentation by human experts. Therefore, there is an urgent need for automated segmentation algorithms to generate high-quality anatomical models with limited data.

Paper Source

This paper is written by Alessandro Fasse, Taylor Newton, Lucy Liang, Uzoma Agbor, Cecelia Rowland, Niels Kuster, Robert Gaunt, Elvira Pirondini, and Esra Neufeld from the Foundation for Research on Information Technologies in Society (IT’IS) and Swiss Federal Institute of Technology (ETH) in Switzerland, as well as the Rehab and Neural Engineering Labs and Center for Neural Basis of Cognition at the University of Pittsburgh, USA. The paper was published in the academic journal “Journal of Neural Engineering” in 2023.

Research Methods

The paper introduces an automated spinal cord MRI image segmentation and model generation pipeline based on Convolutional Neural Networks (CNN), which includes the following steps:

1. Image Preprocessing

   • N4ITK bias field correction is used to improve the uniformity of MR images.
   • An adaptive masking algorithm is used to extract the Region of Interest (ROI), reducing background noise.

2. Data Augmentation

   • Different image variants are generated through affine transformations such as translation, rotation, and scaling to enhance the network’s generalization ability. Each training image can generate 2 to 5 augmented images.

3. Transfer Learning

   • The network weights pre-trained on another dataset are used for initialization to reduce training time and improve performance.

4. Post-processing

   • An automated cleaning procedure is used to identify and retain the main connected component while removing isolated noise regions, ensuring model consistency.

Data Acquisition and Annotation

Spinal cord MRI data from three cats are used: 1. LS1: The lumbosacral region of a cat. 2. LS2: The lumbar and partial sacral region of a cat. 3. CS1: The cervical region of a cat.

Each sample is manually segmented and annotated, providing the ground truth data for training and validating the neural network. The segmentation includes tissues such as cerebrospinal fluid (CSF), dorsal roots, epidural fat, gray matter, ventral roots, and white matter.

Neural Network Architecture

An optimized HardNet architecture is used, combining the HardNet network blocks that perform well with small datasets and the Receptive Field Block (RFB) decoder, forming multiple layers of convolutions and transformations:

• HardNet Blocks: Optimized for GPU memory bandwidth utilization and accelerated convergence by reducing the number of weights.
• Receptive Field Blocks (RFB): Enhance the network’s generalization ability and accuracy on small datasets.

Training Details

• Data Augmentation: Various affine transformations are applied using the Python library Albumentations to increase image diversity.
• Optimization Algorithm: The Adam optimizer is used for backpropagation, with a gradually decreasing learning rate to improve network convergence.

Performance Evaluation

The segmentation performance is evaluated using the following metrics: 1. Jaccard Index: Used to assess the similarity between the network segmentation results and manual annotations. 2. Hausdorff Distance: Used to evaluate the boundary precision of the segmentation results. 3. Activation Function (AF): Used to evaluate the simulated activation of neural fibers.

Key Research Results

1. Segmentation Accuracy: The improved HardNet network demonstrated outstanding segmentation performance, with a Jaccard Index of 0.840 and a Hausdorff Distance of 179 μm, outperforming other common network architectures such as ResNet and VGG.
2. Transfer Learning Effect: Transfer learning significantly improved the network’s generalization ability on new datasets, especially when the anatomical regions differed, such as from the lumbosacral to the cervical region.
3. Automated Cleaning Procedure: The automated cleaning procedure further improved the continuity and consistency of the segmentation, ensuring the anatomical correctness of the model.

Research Significance

The CNN-based automated spinal cord image segmentation pipeline developed in this paper significantly improved segmentation efficiency and accuracy while reducing the need for manual intervention. Compared to traditional manual segmentation, it also demonstrated higher consistency, facilitating the development of personalized SCS treatment plans. Furthermore, this pipeline has the potential for widespread application in other medical image segmentation tasks.

Future Research Directions

• Increase the diversity and quantity of training data to enhance the network’s generalization ability.
• Further validate the pipeline on human cadaver spinal cord MRI data to ensure its clinical feasibility.
• Explore more data augmentation and generative adversarial network (GAN) techniques to synthesize more high-quality training data.

Conclusion

This paper proposed an automated CNN-based spinal cord MRI image segmentation pipeline and improved segmentation accuracy through various image preprocessing and data augmentation techniques. Transfer learning and an automated cleaning procedure further enhanced the model’s usability, laying the foundation for the clinical application of personalized spinal cord stimulation therapy. In future research, further improvements in generalization ability and practical application effects can be achieved through more data and more advanced techniques.