Statistical Features for Emboli Identification Using Clustering Technique

Microembolus signals (MES) detected by transcranial Doppler (TCD) ultrasound are similar to the short duration transient signals. In previous researches, an embolus was tracked by using a supervised technique to discriminate the embolus from the background. However, the classification results were found to be affected by many factors and limited under experimental setup conditions. Therefore, a detection system based on the k-means clustering technique (unsupervised learning) is proposed for emboli detection. In order to verify the proposed technique, the signal data sets are also be computed and compared with SVM classifier. The features selected are the measured embolus-to-blood ratio (MEBR), peak embolus-to-blood ratio (PEBR) and statistical features. Five independent data sets of different transmitted frequency, probe location and different depths are identified to evaluate the feasibility of this new proposed method. The overall result show that k-means is better than SVM in term of robustness aspect. This work also revealed the feasibility of the automatic detection of the features-based emboli in which it is very imperative in assisting the experts to monitor the stroke patients.