Yan Jiun-Lin, Chen Yao-Lian, Chen Moa-Yu, Chen Bo-An, Chang Jiung-Xian, Kao Ching-Chung, Hsieh Meng-Chi, Peng Yi-Ting, Huang Kuan-Chieh, Chen Pin-Yuan
Department of Neurosurgery, Keelung Chang Gung Memorial Hospital, Keelung 204, Taiwan.
School of Traditional Chinese Medicine, College of Medicine, Chang Gung University, Taoyuan 333, Taiwan.
Diagnostics (Basel). 2022 Mar 11;12(3):693. doi: 10.3390/diagnostics12030693.
A midline shift (MLS) is an important clinical indicator for intracranial hemorrhage. In this study, we proposed a robust, fully automatic neural network-based model for the detection of MLS and compared it with MLSs drawn by clinicians; we also evaluated the clinical applications of the fully automatic model. We recruited 300 consecutive non-contrast CT scans consisting of 7269 slices in this study. Six different types of hemorrhage were included. The automatic detection of MLS was based on modified Keypoint R-CNN with keypoint detection followed by training on the ResNet-FPN-50 backbone. The results were further compared with manually drawn outcomes and manually defined keypoint calculations. Clinical parameters, including Glasgow coma scale (GCS), Glasgow outcome scale (GOS), and 30-day mortality, were also analyzed. The mean absolute error for the automatic detection of an MLS was 0.936 mm compared with the ground truth. The interclass correlation was 0.9899 between the automatic method and MLS drawn by different clinicians. There was high sensitivity and specificity in the detection of MLS at 2 mm (91.7%, 80%) and 5 mm (87.5%, 96.7%) and MLSs greater than 10 mm (85.7%, 97.7%). MLS showed a significant association with initial poor GCS and GCS on day 7 and was inversely correlated with poor 30-day GOS (p < 0.001). In conclusion, automatic detection and calculation of MLS can provide an accurate, robust method for MLS measurement that is clinically comparable to the manually drawn method.
中线移位（MLS）是颅内出血的一项重要临床指标。在本研究中，我们提出了一种基于神经网络的强大全自动模型用于检测MLS，并将其与临床医生绘制的MLS进行比较；我们还评估了该全自动模型的临床应用。在本研究中，我们招募了300例连续的非增强CT扫描，共7269层。其中包括六种不同类型的出血。MLS的自动检测基于改进的关键点区域卷积神经网络（Keypoint R-CNN），先进行关键点检测，然后在ResNet-FPN-50骨干网络上进行训练。结果进一步与手动绘制的结果以及手动定义的关键点计算进行比较。还分析了包括格拉斯哥昏迷量表（GCS）、格拉斯哥预后量表（GOS）和30天死亡率在内的临床参数。与真实值相比，MLS自动检测的平均绝对误差为0.936毫米。自动方法与不同临床医生绘制的MLS之间的组内相关系数为0.9899。在检测2毫米（91.7%，80%）、5毫米（87.5%，96.7%）以及大于10毫米（85.7%，97.7%）的MLS时，具有较高的敏感性和特异性。MLS与初始GCS较差以及第7天的GCS显著相关，且与30天GOS较差呈负相关（p < 0.001）。总之，MLS的自动检测和计算可为MLS测量提供一种准确、可靠的方法，在临床上与手动绘制方法相当。
