Abstract
Labeling is an important part of any dataset. Though it can be quite tricky when it comes to labeling ultrasound datasets. As many datasets lack the solid ground-truth labelings such as the ones from the biopsy and blood test result, we have to collaborate with human ultrasound experts on labeling the ultrasound datasets. According to our experience with such an effort, the human diagnosis when tried several times have slight inconsistency (intra-personal inconsistency) and of course the inconsistency can also come from inter-personal observation difference. This research effort is covering this topic and trying to find effective methods to get stable data from mildly inconsistent labelings on ultrasound datasets. Simple averaging may work, but we're going to conduct more solid study on this topic as human labeling is an important method for ultrasound datasets preparation.