Active learning with noise modeling for medical image annotation

Abstract

Active learning is an effective solution to select informative training datasets (examples) from which a pre-defined classifier learns for optimizing its performance. It has been widely applied for information extraction, classification, and filtering. Most existing active learning methods do not consider image noise separately to guide the selection of informative examples, which might lead to sub-optimal annotation. Due to the intrinsic presence of noise in images, large amount of images, and varied imaging modalities, using active learning for medical image annotation is an even more challenging task. In this study, we develop a novel low-rank modeling-based multi-label active learning (LRMMAL) method for effective medical image annotation. Different to those traditional active learning methods, the LRMMAL method innovatively measures image noise and combines it with the measures of example label uncertainty and label correlation into a new sampling process to determine most informative examples for annotation. Experimental results on thoracic CT images and comparisons with other four multi-label active learning methods illustrate the superior performance of the LRMMAL method.