| 李硕,付雅晴,郭冬梅,董洋,刘惠.基于平扫MRI机器学习模型评估兔肝纤维化分期[J].中国介入影像与治疗学,2021,18(7):421-425 |
| 基于平扫MRI机器学习模型评估兔肝纤维化分期 |
| Machine learning model based on plain MRI for evaluation on hepatic fibrosis staging of rabbits |
| 投稿时间：2021-03-01 修订日期：2021-05-30 |
| DOI：10.13929/j.issn.1672-8475.2021.07.009 |
| 中文关键词: 肝 纤维化 磁共振成像 诊断,计算机辅助 纹理分析 |
| 英文关键词:liver fibrosis magnetic resonance imaging diagnosis, computer-assisted texture feature |
| 基金项目: |
|
| 摘要点击次数: 1089 |
| 全文下载次数: 362 |
| 中文摘要: |
| 目的 观察基于平扫MRI的机器学习模型评估肝纤维分期的价值。方法 随机将35只雄性大白兔分入实验组30只、对照组5只。实验组予颈部皮下注射四氯化碳与橄榄油1：1混合溶液，建立肝纤维化模型，每周2次；对照组每周2次颈部皮下注射等量生理盐水；均注射10周。于注药后第5、6、7、10周末分批次采集腹部脂肪抑制T1WI（T1WI-FS）；而后麻醉处死动物并取出肝脏，经HE及Masson染色后依据METAVIR评分系统将肝纤维化分为F0~4期。结合病理结果，依次于各期肝叶勾画感兴趣区（ROI）、提取纹理特征、选择特征；而后采用半监督学习（SSL）与主动学习（AL）相结合算法构建分类器模型，统计五期分类识别及两两分类识别评估肝纤维化分期的准确率。结果 实验组造模过程中2只兔死亡，对照组全部存活。共对35只兔选取180个ROI，病理结果显示F0期32个、F1期37个、F2期33个、F3期54个、F4期24个。五期分类识别中，分类器模型对F0、1、2、3、4期评估准确率分别为87.50%（28/32）、40.54%（15/37）、66.6%（22/33）、64.81%（35/54）及37.50%（9/24），平均准确率60.56%（109/180）。两两分类识别中，分类器模型评估F0 vs F1~4、F0 vs F1~2、F0 vs F3~4、F1~2 vs F3~4、F0~2 vs F3~4的平均准确率分别为95.56%（172/180）、95.10%（97/102）、96.36%（106/110）、68.92%（102/148）及65.56%（118/180）。结论 基于平扫MRI的SSL与AL相结合机器学习模型可有效评估兔模型肝纤维化分期。 |
| 英文摘要: |
| Objective To observe the value of machine learning model based on plain MRI for evaluation on hepatic fibrosis staging of rabbits. Methods A total of 35 male rabbits were randomly divided into experimental group (n=30) and control group (n=5). In experimental group, liver fibrosis models were established by subcutaneous injection of carbon tetrachloride and olive oil 1:1 mixed solution twice a week in the neck, while rabbits in control group were subcutaneously injected with the same amount of normal saline twice a week. Rabbits in both groups were injected for 10 weeks. Abdominal fat suppression T1WI (T1WI-FS) were collected in batches at the end of 5, 6, 7 and 10 weeks after injection. Then the animals were sacrificed, and the livers were harvested. After HE and Masson staining, liver fibrosis was divided into F0-4 stages according to METAVIR scoring system. Combined with pathological results, the regions of interest (ROI) were delineated, the texture features were extracted and features were selected in each stage of liver lobe. An algorithm combining semi-supervised learning and active learning were used to build a classifier model. The accuracies of five-stage classification and pairwise classification in assessment of liver fibrosis staging were counted. Results Two rabbits in experimental group died during modeling, while all rabbits in control group survived. A total of 180 ROI were selected, pathological findings showed the stage of liver fibrosis as F0 (n=32), F1 (n=37), F2 (n=33), F3 (n=54) and F4 (n=24). In five-stage classification recognition, the accuracy of the classifier model for F0, 1, 2, 3 and 4 stages was 87.50% (28/32), 40.54% (15/37), 66.67% (22/33), 64.81% (35/54) and 37.50% (9/24), respectively, the average accuracy was 60.56% (109/180). In pairwise classification, the average accuracy of F0 vs F1-4, F0 vs F1-2, F0 vs F3-4, F1-2 vs F3-4, F0-2 vs F3-4 was 95.56% (172/180), 95.10% (97/102), 96.36% (106/110), 68.92% (102/148) and 65.56% (118/180), respectively. Conclusion The emerging SSL and AL classifier model constructed based on plain MRI could effectively evaluate the stage of liver fibrosis in rabbits. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
