Parkinson's disease (PD) is a typical movement disorder. Clinical scoring scale (ie MDS-UPDRS) is usually used to assess the severity of motor symptoms in PD patients. However, this evaluation method is time-consuming and easily affected by the cognitive differences of the evaluators. Since the outbreak of the new crown pneumonia, remote evaluation of PD patients has become increasingly intense.
To this end, Xiaohua Qian’s research group from the School of Biomedical Engineering of Shanghai Jiaotong University cooperated with Dr. Chencheng Zhang of the functional neurosurgery team of Ruijin Hospital Affiliated to Shanghai Jiaotong University School of Medicine on November 19, 2020 in "IEEE Transactions on Neural Systems" and Rehabilitation Engineering" (TNSRE) journal published an online research article titled "Sparse Adaptive GraphConvolutional Network for Leg Agility Assessment in Parkinson's Disease", which is suitable for automated objective assessment of leg flexibility tasks in MDS-UPDRS based on video Deep learning model.
Leg flexibility task is one of the important parts of MDS-UPDRS. In this test, PD patients need to raise each leg to the ground with the largest amplitude and fastest speed. The evaluator will give an evaluation score based on the patient's speed, range, hesitation and pause, and whether the range is gradually reduced. The range of scores is 0-4. In the research of automatic quantitative evaluation of this task, the existing sensor-based methods have the limitations of intrusiveness, regular calibration and calibration, and the vision-based feature engineering method requires manual feature design in advance.
Therefore, the author developed a sparse adaptive graph convolutional network model to realize the automated evaluation of video-based leg flexibility tasks. Specifically:
1) extracts the joint point sequence of PD patients from the video through the advanced human pose estimation model;
2) proposes the sparse adaptive graph convolution unit (SAGCU) to realize the spatial construction of the human skeleton sequence in the video The model, which adaptively encodes the physical and logical dependence of the human body, then embeds the constraints of the sparsity strategy into the cost function to mine discriminative features, and finally unearths the most important spatial structure relationship in the leg flexibility task ; Z1z
3) introduces the temporal context module (TCM), which constructs the context dependence of the video sequence by calculating the correlation of the time position, and captures the global changes of joints during the execution of the leg flexibility task;
4) develops multi-domain attention Force learning module (MDALM), high-level spatio-temporal features are used to guide low-level features to enhance the salient features in the channel domain, and finally realize the integration of time, space, and channel domains.
Figure 1 The implementation framework of the sparse adaptive graph convolutional network model
This model was comprehensively evaluated on the clinical video data set provided by the Ruijin Hospital Affiliated to Shanghai Jiaotong University School of Medicine, which contains 148 bits 870 videos of the patient’s leg mobility tasks. According to literature survey, this is also the largest data set in the current study of automatic evaluation of PD leg mobility tasks. Quantitative and qualitative evaluation and analysis have proved the validity and reliability of the scheme proposed by the author. The scheme achieves an accuracy rate of 70.34% and an acceptable accuracy rate of 98.97%, which is better than other existing leg flexibility tasks. Evaluation methods (including sensor-based methods).
The non-contact method proposed in this paper provides a new solution for automatic motor function assessment and telemedicine of Parkinson's disease, which is helpful for accurate, continuous and remote monitoring of the disease. Introduction to the research group
Qian Xiaohua
Associate professor and doctoral supervisor at the School of Biomedical Engineering, Shanghai Jiaotong University
Laboratory research interests:
1) image processing and machine learning (deep learning) algorithm development;
2) medical image (video) processing and Analysis, and health big data mining and analysis.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-