Wu An-Rong, Hsieh Sun-Yuan, Chou Hsin-Hung, Lai Cheng-Shih, Hung Jo-Ying, Wang Bow, Tsai Yi-Shan
Department of Computer Science and Engineering, National Cheng Kung University, Tainan, Taiwan.
Department of Computer Science and Engineering, National Chi Nan University, Nantou, Taiwan.
Heliyon. 2025 Jan 8;11(2):e41271. doi: 10.1016/j.heliyon.2024.e41271. eCollection 2025 Jan 30.
Brain midline shift (MLS) indicates the severity of mass effect from intracranial lesions such as traumatic brain injury, stroke, brain tumor, or hematoma. Brain MLS can be used to determine whether patients require emergency surgery and to predict patients' prognosis. Since brain MLS is usually emergent, it must be diagnosed immediately. Therefore, this study presents a computer-aided deep-learning method for detecting MLS, aiming to predict mortality in a prognosis-predicting cohort using brain MLS and clinical in-formation. The brain midline is a 3-dimensional structure, but computed tomography (CT) slices are 2-dimensional which limits brain MLS detection. Here we propose a keypoint detection method to detect brain midline on each CT slice, acquiring brain MLS distance and area in each slice. Combined with clinical information, patient mortality can be predicted using the multilayer perceptron (MLP) model. The accuracy, precision, sensitivity, specificity, and F1-score for slice selection with the proposed model are 0.966, 0.952, 0.991, 0.932, and 0.971, respectively. Both MLS distance and volume were precisely predicted at slice-level and case-level with only the slightest error. The detected midlines were clearly separated into left and right brain with a dice coefficient of 0.98. The accuracy and AUC of the MLP model were both above 0.8. The model detected large brain MLS cases well in the prediction of outcomes in the prognosis-predicting cohort. The method performs well on slice selection and brain MLS detection, and predictions of MLS distance and volume combined with clinical information predicts the patient's prognosis well.
脑中线移位（MLS）表明诸如创伤性脑损伤、中风、脑肿瘤或血肿等颅内病变引起的占位效应的严重程度。脑MLS可用于确定患者是否需要紧急手术，并预测患者的预后。由于脑MLS情况通常很紧急，必须立即进行诊断。因此，本研究提出了一种用于检测MLS的计算机辅助深度学习方法，旨在使用脑MLS和临床信息预测预后队列中的死亡率。脑中线是一个三维结构，但计算机断层扫描（CT）切片是二维的，这限制了脑MLS的检测。在此，我们提出一种关键点检测方法来检测每个CT切片上的脑中线，获取每个切片中的脑MLS距离和面积。结合临床信息，可使用多层感知器（MLP）模型预测患者死亡率。所提模型用于切片选择的准确率、精确率、灵敏度、特异度和F1分数分别为0.966、0.952、0.991、0.932和0.971。在切片水平和病例水平上，MLS距离和体积都得到了精确预测，误差极小。检测到的中线清晰地分为左右脑，骰子系数为0.98。MLP模型的准确率和AUC均高于0.8。在预后预测队列的结果预测中，该模型能很好地检测出大脑MLS较大的病例。该方法在切片选择和脑MLS检测方面表现良好，结合临床信息对MLS距离和体积的预测能很好地预测患者的预后。
