| 本文已被：浏览 192次 下载 122次 |
码上扫一扫！ |
| 基于影像组学分析的椎弓根螺钉钉道超声智能评估 |
| 唐天灵1,马烨波1,杨桓2,叶长青1,孔佑进1,杨卓畅1,周昌1,邵杰2,孟炳堃2,王卓然1,陈建刚1,3*,陈自强2* |
|
|
(1. 华东师范大学上海市多维度信息处理重点实验室, 上海 200241;
2. 海军军医大学(第二军医大学)第一附属医院脊柱外科, 上海 200433;
3. 上海中医药大学康复医学院中医智能康复教育部工程研究中心, 上海 201203
*通信作者) |
|
| 摘要: |
| 目的 基于影像组学分析提出一种针对椎弓根螺钉钉道超声图像的分类方法，以对钉道完整性进行评估。方法 利用4例新鲜尸体的胸腰椎标本预建立50个钉道并获取钉道超声图像，选取2 000张图像（钉道完整与破损的样本各1 000个），采用五折交叉验证的方法将数据集按照4∶1的比例进行随机划分，形成训练集和测试集（分别包含1 600个和400个样本）。首先利用大津阈值法找到感兴趣区的最佳半径，然后用pyradiomics提取组学特征，再采用主成分分析算法和最小绝对收缩和选择算子算法分别进行降维和特征筛选，最后分别使用支持向量机（SVM）、logistic回归、随机森林3种机器学习模型和视觉几何组网络（VGG）、残差网络（ResNet）、转换器模型（Transformer）3种深度学习模型对超声图像进行分类。采用准确度对各模型的分类性能进行评估。结果 在感兴趣区半径为230像素时，SVM模型的分类准确度最高，为96.25%；而VGG模型的准确度只有51.29%，logistic回归、随机森林、ResNet、Transformer模型的准确度分别为85.50%、80.75%、80.17%、75.18%。结论 在对椎弓根螺钉钉道超声图像的分类方面，机器学习模型整体上相较于深度学习模型表现更好，其中SVM模型的分类性能最佳，可用于辅助医师诊断。 |
| 关键词: 椎弓根螺钉置入 超声检查 影像组学 支持向量机 机器学习 人工智能 |
| DOI：10.16781/j.CN31-2187/R.20230560 |
| 投稿时间：2023-10-10修订日期：2024-08-26 |
| 基金项目:国家自然科学基金（82151318），上海市自然科学基金（21ZR1478600），上海市科技计划项目（21Y11902500），2023年地方科学发展基金（XZ202301YD0032C），吉林省科技发展计划项目（20230204094YY），2022年度教育部“春晖计划”合作科研项目，海军军医大学（第二军医大学）“远航”计划. |
|
| Intelligent assessment of pedicle screw canals with ultrasound based on radiomics analysis |
| TANG Tianling1,MA Yebo1,YANG Huan2,YE Changqing1,KONG Youjin1,YANG Zhuochang1,ZHOU Chang1,SHAO Jie2,MENG Bingkun2,WANG Zhuoran1,CHEN Jiangang1,3*,CHEN Ziqiang2* |
(1. Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University, Shanghai 200241, China;
2. Department of Spinal Surgery, The First Affiliated Hospital of Naval Medical University (Second Military Medical University), Shanghai 200433, China;
3. Engineering Research Center of Traditional Chinese Medicine Intelligent Rehabilitation, Ministry of Education, Shanghai University of Traditional Chinese Medicine, Shanghai 201203, China
*Corresponding authors) |
| Abstract: |
| Objective To propose a classification method for ultrasound images of pedicle screw canals based on radiomics analysis, and to evaluate the integrity of the screw canal. Methods With thoracolumbar spine specimens from 4 fresh cadavers, 50 pedicle screw canals were pre-established and ultrasound images of the canals were acquired. A total of 2 000 images (1 000 intact and 1 000 damaged canal samples) were selected. The dataset was randomly divided in a 4∶1 ratio using 5-fold cross-validation to form training and testing sets (consisting of 1 600 and 400 samples, respectively). Firstly, the optimal radius of the region of interest was identified using the Otsu’s thresholding method, followed by feature extraction using pyradiomics. Principal component analysis and the least absolute shrinkage and selection operator algorithm were employed for dimensionality reduction and feature selection, respectively. Subsequently, 3 machine learning models (support vector machine ［SVM］, logistic regression, and random forest) and 3 deep learning models (visual geometry group ［VGG］, ResNet, and Transformer) were used to classify the ultrasound images. The performance of each model was evaluated using accuracy. Results With a region of interest radius of 230 pixels, the SVM model achieved the highest classification accuracy of 96.25%. The accuracy of the VGG model was only 51.29%, while the accuracies of the logistic regression, random forest, ResNet, and Transformer models were 85.50%, 80.75%, 80.17%, and 75.18%, respectively. Conclusion For ultrasound images of pedicle screw canals, the machine learning model performs better than the deep learning model as a whole, and the SVM model has the best classification performance, which can be used to assist physicians in diagnosis. |
| Key words: pedicle screw implantation ultrasonography radiomics support vector machine machine learning artifical intelligence |
