Effective Spinal Cord Diagnosis using a Hybrid Model of 3D U-Net and CNN

Abstract

Introduction: Spinal cord injuries (SCI) can have severe, life-changing implications, therefore early and precise diagnosis is critical. While CT imaging is useful, hand interpretation is slow and error-prone, emphasizing the importance of automated technologies. Our paper proposes a deep learning-based technique that uses a 3D CNN to segment the spinal cord and detect lesions in CT images. By training on a huge dataset, our model aspires to match or outperform human experts in accuracy while streamlining clinical workflows. Finally, we want to improve patient outcomes by incorporating deep learning into routine spinal injury examinations.

Objectives: The goal of this research is to create a deep learning-based system that can accurately segment the spinal cord and identify injuries in CT images. Utilizing a 3D CNN model that has been trained on a sizable annotated dataset, our goal is to improve diagnostic efficiency and accuracy. Our objectives are to assist radiologists, lower diagnostic mistakes, and enhance clinical results for patients with spinal cord injuries.

Methods: Early and precise diagnosis is essential for spinal cord injuries (SCI), which can result in paralysis and lifelong disability. Although CT imaging is often utilized, manual interpretation is laborious and prone to mistakes. In order to solve this, we suggest a deep learning-based system for spinal cord segmentation and damage detection that makes use of a 3D CNN. To increase accuracy and help radiologists make quicker, more accurate diagnoses, our model is trained on annotated CT images. Our goal is to improve patient outcomes and healthcare procedures by automating this process.

Results: This study divides methods for detecting spinal cord injuries into three categories: distance-based, qualitative, and quantitative. With a mean of 97% in volumetric similarity and over 94% similarity with ground truth, the suggested hybrid CNN and 3D U-Net model performs better than the rest in segmentation accuracy. Its higher performance is confirmed by traditional measures like sensitivity, specificity, and MCC, which shows a significant connection between forecasts and ground truth.

Conclusions: This study divides models for detecting spinal cord injuries (SCI) into three categories: distance-based, qualitative, and quantitative. With over 97% similarity and almost 0% Global Consistency Error (GCE), the suggested hybrid CNN and 3D U-Net model performs better in segmentation accuracy than conventional techniques. Superior sensitivity and specificity are confirmed by quantitative assessment utilizing measures such as the Dice coefficient, Jaccard index, and Matthews correlation. Its accuracy is further confirmed by distance-based metrics like Hausdorff Distance (HD) and Mean Surface Distance (MSD).