基于多尺度特征提取和时间分割的梅杰综合征检测-《中国医学物理学杂志》

文章信息/Info

Title:: Detection of Meige’s syndrome based on multi-scale feature extraction and temporal segmentation

作者:: 李碧草1; 衣本泽1; 王贝2; 刘志涛3; 郭旭伟4; 王岩1; 1.中原工学院信息与通信工程学院，河南郑州 450007；2.中原工学院校医院，河南郑州 451191；3.河南省直第三人民医院梅杰诊疗中心，河南郑州 450018；4.河南科技大学第一附属医院儿科，河南洛阳 471000

Author(s):: LI Bicao1; YI Benze1; WANG Bei2; LIU Zhitao3; GUO Xuwei4; WANG Yan1; 1. School of Information and Communication Engineering, Zhongyuan University of Technology, Zhengzhou 450007, China; 2. UniversityInfirmary, Zhongyuan University of Technology, Zhengzhou 451191, China; 3. Diagnosis and Treatment Center of Meige’s Syndrome,the Third People’s Hospital of He’nan Province, Zhengzhou 450018, China; 4. Pediatric Department, the First Affiliated Hospital ofHe’nan University of Science and Technology, Luoyang 471000, China

Keywords:: Meige’s syndrome; temporal action detection; multi-scale feature extraction; temporal segmentation

摘要:: 梅杰综合征的诊断主要依赖医生的临床评估，由于该疾病的症状与其他神经性疾病有相似之处且确诊较为复杂，因此诊断过程对医生和患者相当关键。本研究采集31名梅杰综合征患者的发病视频，建立梅杰综合征检测数据集，开发出一种应用于未修剪视频的梅杰综合征自动检测系统（MS-Net）。首先，利用RetinaNet和UNet3+构建时间检测分支和时间分割分支来进行多尺度特征提取和时间分割；其次，通过时间检测分支和时间分割分支解码分别生成检测窗口的概率向量和每帧发病的概率；最后，使用多层感知器处理两分支的概率预测，为每个窗口生成一个更加准确的概率。通过使用额外的损失函数和数据增强等技术优化模型性能，并使用临床医生可解释特征进行操作。MS-Net可以辅助诊断梅杰综合征，提高早期诊断的准确性、便捷性和效率。与其他先进的网络进行比较，结果表明MS-Net在使用临床实践中所需的可解释特征的同时在平均精度方面也取得相当的性能。

Abstract:: Abstract: The diagnosis of Meige’s syndrome predominantly relies on the clinical assessment by physicians. Given thecomplexity and similarity of its symptoms to other neurological disorders, the diagnosis is crucial for both doctors andpatients. Herein a detection dataset for Meige’s syndrome is compiled from video recordings of 31 patients, and an automateddiagnostic system for Meige’s syndrome (MS-Net) applicable to untrimmed videos is developed. The system utilizesRetinaNet and UNet3+ to construct temporal detection and segmentation branches for multi-scale feature extraction andtemporal segmentation, obtains probability vectors for detection windows and the probability of disease onset per frame viathe decoding of temporal detection and segmentation branches, and finally generates a refined probability for each windowby processing the probability predictions from both branches using a multi-layer perceptron. The model performance isoptimized using additional loss functions and data augmentation techniques, operating on features interpretable by clinicalphysicians. MS-Net can assist in the diagnosis of Meige’s syndrome, improving the accuracy, convenience, and efficiency ofthe early diagnosis. The comparison of MS-Net with other state-of-the-art networks indicates that MS-Net achievescomparable performance in terms of average precision while utilizing interpretable features required in clinical practice.