Accurate tumor identification is crucial for diagnosing and treating various diseases. Nevertheless, the limited availability of expert pathologists delays reliable and timely tumor identification. Crowdsourcing can assist by taking advantage of the collective intelligence of crowdworkers through consensus-based opinion aggregation. However, the open problem of training crowdworkers rapidly for doing complex tasks poses a significant challenge, currently yielding inaccurate results. To improve the performance of crowdworkers, we present a redesign of the training strategy by addressing the errors crowdworkers face frequently. By identifying error patterns through a study, we optimize the design of the training strategy for an exemplary tumor identification crowdsourcing task. We conduct a comparative analysis between a baseline version of the training strategy and an optimized version based on identified error patterns. Our findings demonstrate that optimizing the training strategy significantly reduces annotation mistakes during the crowdsourced tumor identification process, attributable to the increase of retention. Moreover, it provides noticeable improvements in the performance of annotation of correct tumor regions. This research contributes to the field by testing the effectiveness of training strategy optimization in crowdsourcing tasks, specifically for tumor annotation. Addressing crowdworkers' training needs and leveraging their collective intelligence, our approach enhances tumor identification's reliability, providing alternatives for healthcare decision-making.
