Semi-supervised annotation of Transcranial Doppler ultrasound micro-embolic data

Transcranial Doppler (TCD) is a non-invasive ultrasound monitoring technique allowing real time measurements of the blood flow velocity mainly in the Middle Cerebral Artery. It is commonly used to monitor patients with stroke risk by detecting micro-emboli. This technique generates a considerable amount of data whose annotation is expensive and time-consuming. We propose a semi-supervised learning method to annotate a dataset of micro-embolic data with reduced requirements of manual annotation. First, we start by extracting features from TCD spectrograms in an unsupervised manner using an Auto-encoder. Then, we project those features in a 2D space using the t-SNE algorithm. Afterwards, the dataset is partially annotated by an expert based on the 2D projection and the overall data information. Finally, the labels of the annotated samples are propagated to a part of the unlabeled samples using a K-nearest-neighbors (KNN) strategy. We evaluate our annotation method through the annotation error and the percentage of labeled samples with respect to the unlabeled samples. Our results show that we are able to annotate all the unlabeled samples with an annotation error of 12 % in less than 1 second against around 2 hours for a human expert. This represents a time saving of a factor of 10
⁵
, showing the interest of our method.