CrowdAttention: An Attention Based Framework to Classify Crowdsourced Data in Medical Scenarios

Supervised learning models in healthcare and other domains heavily depend on high-quality, labeled data. However, acquiring expert-verified labels (i.e., the gold standard) is often impractical due to cost, time, and subjectivity. Crowdsourcing offers a scalable alternative by collecting labels from multiple non-expert annotators; however, it introduces label noise due to the heterogeneity of annotators. In this work, we propose CrowdAttention, a novel end-to-end deep learning framework that jointly models classification and annotator reliability using a cross-attention mechanism. The architecture consists of two coupled networks: a classification network that estimates the latent true label, and a crowd network that assigns instance-dependent reliability scores to each annotator's label based on its alignment with the model's current prediction. We demonstrate the effectiveness of our approach on both synthetic and real-world datasets, showing improved accuracy and robustness compared to state-of-the-art multi-annotator learning methods.