Transcranial Doppler (TCD) is an ultrasound monitoring method that enables real-time measurements of blood flow velocity, primarily in the Middle Cerebral Artery. Its common application lies in monitoring patients at risk of stroke by effectively detecting micro-emboli. This is done through the detection of high intensity transient signals (HITS), which can be categorized between artifacts, gaseous emboli, and solid emboli. State-of-the-art methods for HITS classification are not able to capture the uncertainty of HITS soft-annotation, nor the noise in their soft-labels (soft-noise), both coming from the expert annotation doubt. To better handle this, we propose to train deep learning models using soft labels cost functions, such as the soft cross entropy and the Jensen-Shanon divergence (JSD), which directly approximates the soft distribution of the input samples instead of a hard label proxy. We evaluate the robustness of our approach against symmetrical soft label noise in terms of final hard classification (using the Matthews correlation coefficient, MCC) and human expert uncertainty capturing (using the Hellinger distance). The obtained models trained with JSD soft-labels were robust against soft-noise with improvements of up to 24% in terms of MCC. At last, these models were able to better capture the human expert uncertainty of the true labels, achieving Hellinger distance improvements up to 0.10 (relative gap of 16%).