Classification of Red Blood Cell Aggregation with Hyper Spectral Analysis of Ultrasonic Radiofrequency Echo Signals

Abstract

In this paper, red blood cell aggregation classification based on hyper spectral analysis of ultrasonic radiofrequency (RF) echo signals is proposed. Firstly, Morlet wavelet is applied to the sub-band decomposition of ultrasonic RF echo signals. Then, five statistical features including mean, variance, median, kurtosis and root mean square of each sub-band are calculated to form the feature vectors. 18 kinds of blood with different red blood cell concentration-aggregation are taken as samples, then multi-frame ultrasonic RF echo signals are collected using ultrasonic linear array probe. The region of interest (ROI) is selected from the B-mode image of a certain frame. 20 subbands are obtained by the hyper spectral analysis of each line of ultrasonic RF echo signals in the ROI. Five statistical features of each sub-band are calculated, and then the feature vectors are obtained after local normalization. Finally, support vector machine (SVM) and random forest classifiers are used to classify the feature vectors respectively. The overall average classification accuracy of SVM is $91.43\pm 6.17\%$, and the overall average classification accuracy of random forest classifier is 96.19 ± 4.28 %.