A Noise Label Correction Architecture for Medical Image Segmentation
Hui Wang, Sos С. Agaian, Desheng Liu · 2024
The efficacy of deep learning in medical image segmentation is heavily reliant on large datasets with precise annotations. Yet, acquiring such data is costly and labor-intensive, and the annotations are often plagued by errors, particularly for subtle lesions. These inaccuracies can lead to artificial intelligence (AI) misdiagnoses in early-stage liver cancer. However, discarding noisy data will not only decrease the amount of data, but also degrade the performance of learning model. The available label correction methods may have label correction errors in the process of correcting noisy labels. To combat the limitations of noisy data and flawed label correction methods, we present the Confident Teacher Correction Network (CTCNet). CTCNet replaces unreliable labels with ‘soft’ labels that are refined during training to minimize noise. Our novel loss function mitigates the impact of high-variance predictions, reducing the uncertainty of predictions. CTCNet consists of two steps: (1) using the proposed shallow denoising subtraction network to remove background noise and learning high frequency information of small focus areas and (2) correcting the learning of noisy labels using a new loss function. We evaluated CTCNet on two public datasets of noise labels with different erasure rates. The results of the experiments showed that CTCNet had demonstrated superior performance over existing state-of-the-art methods. In particular, CTCNet does not require prior knowledge of noise characteristics, making it a robust and versatile solution. We can use CTCNet to assist the correction of the doctor's annotation, which can provide physicians and researchers with more reliable image annotation.