| 孟凡盛,田占宇,龚伟,熊亮,蒋海壮,于丽娟.18F-D3FSP PET/CT机器学习模型评估主观认知功能下降[J].中国医学影像技术,2025,41(4):573~577 |
| 18F-D3FSP PET/CT机器学习模型评估主观认知功能下降 |
| 18F-D3FSP PET/CT machine learning models for evaluating subjective cognitive decline |
| 投稿时间：2024-09-23 修订日期：2025-03-10 |
| DOI：10.13929/j.issn.1003-3289.2025.04.014 |
| 中文关键词: 认知障碍 机器学习 体层摄影术,X线计算机 正电子发射断层显像 |
| 英文关键词:cognition disorders machine learning tomography, X-ray computed positron-emission tomography |
| 基金项目:海南省自然科学基金青年基金项目(822QN483)。 |
|
| 摘要点击次数: 234 |
| 全文下载次数: 75 |
| 中文摘要: |
| 目的 观察18F-D3FSP PET/CT机器学习(ML)模型用于评估主观认知功能下降(SCD)的价值。方法 于中国认知下降纵向研究队列(SILCODE)中选取32例接受18F-D3FSP PET/CT检查的SCD患者(SCD组)及16名健康志愿者(对照组);按7∶3比例将其分为训练集(n=34)及测试集(n=14),基于汉密尔顿焦虑量表(HAMA)评分及组间差异有统计学意义的脑区标准摄取值比值(SUVR)分别构建支持向量机(SVM)、随机森林(RF)及逻辑回归(LR)模型评估SCD;利用格式转化对PET/CT数据进行扩增,并按8∶2比例划分训练集(含8 775幅CT及1 833幅PET图像)及测试集(含2 025幅CT及423幅PET图像),分别基于CT及PET构建VGG16模型用于评估SCD。结果 SVM、RF及LR模型评估训练集SCD的受试者工作特征曲线下面积(AUC)均为1.000,在测试集分别为0.863、0.872及1.000;LR模型存在过拟合,RF模型效能更优。基于CT及基于PET的VGG16模型评估训练集SCD的准确率分别于第175次及第150次迭代后趋于稳定,最高分别为67.11%及65.35%;其在测试集的准确率分别于第165次及第145次迭代后趋于稳定,最高分别为62.43%及59.16%。结论 18F-D3FSP PET/CT ML模型可用于评估SCD。 |
| 英文摘要: |
| Objective To observe the value of 18F-D3FSP PET/CT machine learning (ML) models for evaluating subjective cognitive decline (SCD). Methods Thirty-two SCD patients (SCD group) and 16 healthy volunteers (control group) who received 18F-D3FSP PET/CT were selected from Sino Longitudinal Study on Cognitive Decline (SILCODE) and divided into training set (n=34) and test set (n=14) at a ratio of 7 ∶ 3. Support vector machine (SVM), random forest (RF) and logistic regression (LR) models were constructed based on Hamilton anxiety scale (HAMA) and standard uptake value ratio (SUVR) of brain regions being significantly different between groups to evaluate SCD. Then PET/CT data were amplified by format conversion and divided into training set (including 8 775 CT images and 1 833 PET images) and test set (including 2 025 CT images and 423 PET images) at the ratio of 8∶2. VGG16 models were constructed based on CT and PET images to evaluate SCD, respectively. Results The area under the receiver operating characteristic curve (AUC) of SVM, RF and LR model for evaluating SCD in training set was all 1.000, while was 0.863, 0.872 and 1.000 in test set, respectively. LR model was overfitting, and RF model had better performance. The accuracy of VGG16 model for evaluating SCD based on CT and PET images tended to be stable after the 175th and 150th iterations, with the highest accuracy of 67.11% and 65.35% in training set, which tended to be stable after the 165th and 145th iterations, with the highest accuracy of 62.43% and 59.16% in test set, respectively. Conclusion 18F-D3FSP PET/CT ML models could be used to evaluate SCD. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
