Region-scalable fitting-assisted medical image segmentation with noisy labels

Abstract

To aid in diagnosis and analysis, deep learning has been applied to medical image analysis by many researchers. Most existing methods train models using many images with precise labels. However, medical images are inherently complex and noisy, making it difficult for medical image segmentation to obtain many accurate labels. That means the study of learning with noisy labels is particularly important. In order to make use of all the samples, some methods correct the noisy labels during the training process. However, in the correction process, a label correction error may occur according to the set policy. Therefore, we propose a method that incorporates image features based on region-scalable fitting (RSF) to assist in the label correction process, called RSF-assisted. We tested the performance of our RSF-assisted method on two learning frameworks of learning with noisy labels based on segmentation: mean-teacher-assisted confident learning (MTCL) and adaptive early-learning correction (ADELE). The experimental results indicate that the performance is improved by using RSF-assisted. To be specific, compared with the original methods, the mean intersection over union (mIOU) is improved by 2.41% after RSF-assisted on thoracic organs (SegTHOR) dataset. Similarly, the 95% Hausdorff distance (95HD) is reduced by up to 2.96 mm after RSF-assisted based on left atrium (LA) segmentation dataset.