| 黄婷婷,王卓琛,赵鑫,杨凯华,张晗宇,李曼,邢威,张刚.自引导注意力网络检测脑室周围白质损伤患儿脑瘫相关责任病灶[J].中国医学影像技术,2025,41(5):723~728 |
| 自引导注意力网络检测脑室周围白质损伤患儿脑瘫相关责任病灶 |
| Self-guided attention network for detecting responsible lesions related to cerebral palsy in children with periventricular white matter injury |
| 投稿时间：2024-08-06 修订日期：2025-02-03 |
| DOI：10.13929/j.issn.1003-3289.2025.05.006 |
| 中文关键词: 脑损伤 脑性瘫痪 脑室周围白质损伤 磁共振成像 人工智能 |
| 英文关键词:brain injuries cerebral palsy periventricular white matter injury magnetic resonance imaging artificial intelligence |
| 基金项目:国家自然科学基金(82204933)、河南省卫健委国家中医药临床研究基地科研专项(2022JDZX094)。 |
|
| 摘要点击次数: 56 |
| 全文下载次数: 13 |
| 中文摘要: |
| 目的 观察自引导注意力网络检测脑室周围白质损伤(PVWMI)患儿脑瘫(CP)相关责任病灶的效能。方法 回顾性纳入383例PVWMI患儿,将其分为CP组(n=243)与非CP组(n=140);另以214例颅脑MRI无明显异常患儿为对照组。基于颅脑T1WI分别于半卵圆中心、内囊后肢、大脑脚及丘脑勾画 ROI,于T2WI中标注其内CP相关责任病灶,将二者配准作为输入网络。采用ResNet34网络、结合注意力及自引导网络训练网络并检测PVWMI患儿CP相关责任病灶,评估其检测关键解剖结构中病灶的效能以筛选最优者,观察其分割关键解剖结构的效能。结果 自引导注意力网络为最优网络,其检出病灶的曲线下面积为0.794~0.914,分割关键解剖结构的戴斯相似系数为0.702~0.764。结论 自引导注意力网络可用于初步检测PVWMI患儿CP相关责任病灶。 |
| 英文摘要: |
| Objective To observe the efficacy of self-guided attention network for detecting responsible lesions related to cerebral palsy (CP) in children with periventricular white matter injury (PVWMI). Methods Totally 383 children with PVWMI were retrospectively enrolled and divided into CP group (n=243) and non-CP group (n=140), while 214 children without obvious brain abnormality on brain MRI were taken as control group. ROI of 4 key anatomical structures related to CP, i.e. centrum semiovale, posterior limb of internal capsule, cerebral peduncle and thalamus were delineated on T1WI, while responsible lesions related to CP within the key anatomical structures were labeled on T2WI, and the images were then registrated and used as input of the networks. ResNet34 network was adopted combined with attention and self-guided networks to train the network for detecting responsible lesions related to CP in children with PVWMI, and their efficacies were evaluated. The optimal network was screened, and its efficacy for segmenting the key anatomical structures was evaluated. Results Self-guided attention network was the optimal network, its area under the curve (AUC) for detecting lesions was 0.794—0.914, and the Dice similarity coefficient for segmenting the key anatomical structures was 0.702—0.764. Conclusion Self-guided attention network could be used for preliminarily detecting responsible lesions related to CP in children with PVWMI. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
