Deep Self-Cleansing for Medical Image Segmentation with Noisy Labels

TL;DR

This research presents a novel approach to segmentation that combines supervised deep learning with reinforcement learning to solve the challenge of integrating image classification and annotation into a discrete-time model.

Abstract

Medical image segmentation is crucial in the field of medical imaging, aiding in disease diagnosis and surgical planning. Most established segmentation methods rely on supervised deep learning, in which clean and precise labels are essential for supervision and significantly impact the performance of models. However, manually delineated labels often contain noise, such as missing labels and inaccurate boundary delineation, which can hinder networks from correctly modeling target characteristics. In this paper, we propose a deep self-cleansing segmentation framework that can preserve clean labels while cleansing noisy ones in the training phase. To achieve this, we devise a gaussian mixture model-based label filtering module that distinguishes noisy labels from clean labels. Additionally, we develop a label cleansing module to generate pseudo low-noise labels for identified noisy samples. The preserved clean labels and pseudo-labels are then used jointly to supervise the network. Validated on a clinical liver tumor dataset and a public cardiac diagnosis dataset, our method can effectively suppress the interference from noisy labels and achieve prominent segmentation performance.

Figures (7)

• Figure 1: Typical types of label noise. Take CT slices with liver tumors as an example, (a) illustrates a missing label that radiologists omitted the tumor region. (b) illustrates the boundary noises around the tumor region.
• Figure 2: Schematic view of the deep self-cleansing network, which filters out the Noisy labels through LFM and cleanses them through LCM interatively.
• Figure 3: Illustration of grid-based representation maps.
• Figure 4: Illustration of the training pipeline.
• Figure 5: Qualitative results of different methods on the HCC dataset (B-model, $\alpha_1=70\%$, $\beta_1=10$), where liver tumor regions are highlighted in red, and discrepancies between network-produced results and ground truths are highlighted in yellow.