(C-96) A Miniature M-Mode Ultrasound based on Time Delay Spectrometry for Hand Grasps Classification

Introduction:: Introduction: Sonomyography (SMG), an emerging modality that uses ultrasound to detect muscle deformation, has shown promise as a method for correctly classifying intended hand grasps of those with upper limb loss from functional activity of muscle. Conventional pulse-echo ultrasound systems require dedicated high voltage pulser receivers and high frequency digitizers. We have developed a novel imaging technique based on time-delay spectrometry (TDS) that utilizes low voltage frequency encoded transmissions and audio frequency digitizers that are more suitable for miniaturization. The objective of this study was investigating the reliability of TDS to be used in hand grasps classification.


Materials and Methods:: Materials and Methods: In this study, we employed TDS system to obtain 4-channel M-mode images of the in vivo muscle interfaces of ten subjects. The participants were asked to perform 5 different grasps including Rest, Power, Tip, Point, and Key which is shown in Figure 1. The subjects performed the grasps while keeping their hands in a location near to their body at 90 degrees arm angle and did different grasps for 5 seconds statically. We used Principal Component Analysis (PCA), which can explain 70% of the variance in the data matrix. After that, statistical features including Maximum, Minimum, Standard Deviation, Mean, Skewness, Kurtosis, Z score and sum of absolute differences were extracted. We classified the extracted features using Decision Tree Classifier (DTC) and K-Nearest Neighborhood (KNN) as well as Linear Discriminant Analysis (LDA). Accuracy was calculated to present the reliability of using TDS system to capture M-Mode images for purpose of classification hand grasps applications.

< ![if !vml] >< ![endif] >

Figure 1. Different hand grasps for data collection.


Results, Conclusions, and Discussions:: Results and Discussion: The average classification accuracy for each classifier is shown in Table 1, which illustrates classification accuracy for DTC and KNN which is 98.46% and 92.31% respectively and more information on the f1 Score and Recall metrics. The LDA algorithm did not produce acceptable results. The DTC algorithm performed well. We interpret these results to indicate that a nonparametric algorithm like DTC is more robust to variability compared to LDA. This study demonstrates the possibility of classifying different hand grasps by capturing M-Mode ultrasound images using TDS System which provides us reliable results for different classifiers.

Table 1. Classification metrics for 2 applied methods.

Metrics Classifier	Accuracy	F1 Score	Recall
Decision Tree Classifier	98.46%	96.48%	98.45%
K-Nearest Neighbor	92.31%	92.30%	92.16%

Acknowledgements (Optional): :

References (Optional): :