Application of deep learning models for detection of subdural hematoma: a systematic review and meta-analysis
Application of deep learning models for detection of subdural hematoma: a systematic review and meta-analysis
复制标题
深度学习模型在硬膜下血肿检测中的应用：系统评价和荟萃分析
DOI：
10.1136/jnis-2023-020218
复制
发表时间：
2023
影响因子：
4.8
通讯作者：
Agarwal S
中科院分区：
文献类型：
--
作者：
Agarwal S
我们感兴趣地阅读了Abdollahifard等人最近发表的文章。1人工智能（AI）检测严重异常（如硬膜下血肿）的能力在全球范围内都是相关的，不仅是因为作者讨论的诊断准确性的潜在收益，而且还因为成像分诊。[2]作者花了相当大的努力筛选了9485篇摘要。他们报告了22种深度学习算法检测硬膜下血肿的准确性，并将其中11项研究纳入荟萃分析。然而，我们确实担心哪些研究被纳入和偏倚评估的风险，我们最担心的是作者错误地混淆了检测准确性和分割准确性。检测是AI模型能够将检查作为输入并给出（通常）二进制输出的能力，无论检查是否包含异常。分割是AI模型将检查作为输入并返回检查中异常体素的逐体素图的能力。分割研究中的假阳性是指预测异常图中包含多少正常体素，而不是错误预测有多少健康患者患有硬膜下血肿。分割和检测AI模型具有不同的功能，这些准确性无法比较，因此荟萃分析中至少有两项研究报告了分割准确性而不是检查级检测准确性。3 4
We read with interest the recent article by Abdollahifard et al. 1 The ability of artificial intelligence (AI) to detect critical abnormalities, such as subdural hematomas, is relevant globally, not only for the potential gains in diagnostic accuracy as the authors discuss, but also for imaging triage. 2 The authors have undertaken a considerable effort to screen 9485 abstracts. They report the accuracy of 22 deep learning algorithms that detect subdural hematomas and included 11 of these studies in a meta-analysis. We do, however, have concerns regarding which studies have been included and the risk of bias assessment.Our biggest concern was that the authors have incorrectly conflated detection accuracy and segmentation accuracy. Detection is the ability of an AI model to be able to take an examination as an input and give a (usually) binary output whether the examination contains an abnormality. Segmentation is the ability of an AI model to take an examination as input and return a voxel-by-voxel map of abnormal voxels within an examination. False positives in a segmentation study refer to how many normal voxels are included in a predicted abnormality map, rather than how many healthy patients are erroneously predicted to have a subdural hematoma. Segmentation and detection AI models have different functions, and these accuracies cannot be compared, so it was concerning that at least two studies included in the meta-analysis reported segmentation accuracy rather than examination-level detection accuracy. 3 4
影响因子：
4.4
作者：
Cho, Junghwan;Park, Ki-Su;Park, Sinyoul
通讯作者：
Park, Sinyoul
用于诊断 CT 扫描中颅内出血的 ResNet-18 和 DenseNet-121 模型的迁移学习。
影响因子：
3.1
作者：
Qi Zhou;Wenjie Zhu;Fuchen Li;M. Yuan;Linfeng Zheng;Xu Liu
通讯作者：
Xu Liu
影响因子：
2
作者：
Kellogg, Ryan T.;Vargas, Jan;Levitt, Michael
通讯作者：
Levitt, Michael