A0981
Title: Imaging data classification with crowdsourced noisy annotations
Authors: Guanqun Cao - Michigan State University (United States) [presenting]
Abstract: In typical label collection processes for imaging data, multiple annotators provide subjective, noisy estimates of the truth, influenced by their varying skill levels and biases. Treating these noisy labels as ground truth can significantly hinder the accuracy of learning algorithms, particularly in cases of strong disagreement. To address this challenge, the annotated functional deep neural network (afDNN) method is proposed, a novel data mining and classification framework for noisy annotated imaging data by modeling it as noisy labeled functional data. This architecture leverages a sparse deep neural network to jointly model individual annotator behavior and infer the underlying true label distribution from noisy observations. To enhance performance, a regularization term is incorporated into the loss function, encouraging convergence toward the true annotator confusion matrix. Additionally, the convergence rates of the misclassification risk functions have been derived for both densely and sparsely observed imaging data. The effectiveness of afDNN is demonstrated through simulations and applications of two real biomedical imaging datasets.
