基于深度学习的椎间孔狭窄自动多分级研究

PDF下载HTML

Title:

Deep learning based automatic multi-classification algorithm for intervertebral foraminal stenosis

作者:

洪雁飞^1,2, 魏本征^1,2, 刘川², 韩忠义^1,2, 李天阳^1,2

1. 山东中医药大学 理工学院, 山东 济南 250355;
2. 山东中医药大学 计算医学实验室, 山东 济南 250355

Author(s):

HONG Yanfei^1,2, WEI Benzheng^1,2, LIU Chuan², HAN Zhongyi^1,2, LI Tianyang^1,2

1. College of Science and Technology, Shandong University of Traditional Chinese Medicine, Ji’nan 250355, China;
2. Computational Medicine Lab, Shandong University of Traditional Chinese Medicine, Ji’nan 250355, China

关键词:

椎间孔狭窄; 自动分级; 机器学习; 深度学习; 特征提取; 监督训练; 迁移学习; 过拟合

Keywords:

intervertebral foraminal stenosis; automatic grade; machine learning; deep learning; feature extraction; supervised training; transfer learning; over fitting

分类号:

TP311

DOI:

10.11992/tis.201806015

摘要:

椎间孔狭窄症的术前定性分级诊断对临床医生治疗策略的制定和患者健康恢复至关重要，但目前该方面临床上仍然存在很多问题，并且缺乏相关的研究和行之有效的方法用于辅助临床医生诊断。因此，为提高计算机辅助椎间孔狭窄症诊断准确率以及医生工作效率，本文提出一种基于深度学习的椎间孔狭窄图像自动分级算法。从人体矢状切脊柱核磁共振图像中提取脊柱椎间孔图像，并做图像预处理；设计一种监督式深度卷积神经网络模型，用于实现脊柱椎间孔图像数据集的自动多分级；利用迁移学习方法，解决深度学习算法在小样本数据集上的过拟合问题。实验结果表明，本文算法在脊柱椎间孔图像数据集上的分类精确度可达到87.5%以上，且其具有良好的鲁棒性和泛化能力。

Abstract:

Preoperative qualitative diagnosis of intervertebral foraminal stenosis is essential for the formulation of clinician treatment strategies and patients’ health recovery. However, there are still many clinical challenges in this aspect and a lack of relevant research and proven methods to assist clinicians in diagnosis. Therefore, a deep learning-based automatic classification algorithm is proposed in this study to improve the diagnosis accuracy and the efficiency. First, we extracted the spinal foramen images from the sagittal spine MRI image, and then these images were preprocessed. Second, a supervised deep convolutional neural network model was designed to achieve automatic multi-classification for the datasets of the intervertebral foraminal stenosis. Finally, we used the transfer learning to optimize the overfitting problem of the deep learning algorithm in the small sample dataset. The experimental results show that the classification accuracy of this algorithm on the dataset of spinal foramen was 87.5%, and it has good robustness and generalization performance.