Bayesian Optimisation Active-Learning Semi-Supervised Classification for an Automatic Microembolus-Detection System

Embolic stroke is a top life-threatening disease for which an early diagnosis is favourable. Commonly, the diagnosis of the microembolus in circulating blood flow is performed by experts manually using a transcranial Doppler ultrasound device. However, correctly identifying microembolus signals over a long period of time is challenging. Therefore, an automatic microembolus-classification system that can have expert human labelling capabilities is required in classifying microembolus and artifact signals. In this work, we use Bayesian optimisation incorporated into the standard active-learning approach. This technique further optimises and corrects bad labelling from previous predictions, thereby helping reduce the number of bad examples and increasing the overall classification performance. In feature selection, a feature score is introduced, and the overall score is fine tuned using neighbourhood component analysis to obtain the most significant features for the training data. The proposed system performance is then compared with standard active-learning, self-training, and baseline systems. The results proved that the proposed system can classify unlabelled data with an automated labelling process without the presence of a human expert. The experimental results of the Bayesian optimisation active-learning system produced performance-evaluation parameters of 87.34%, 90.96%, and 80.99% for accuracy, sensitivity, and specificity, respectively.