Imbalanced Medical Image Segmentation With Pixel-Dependent Noisy Labels.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Accurate medical image segmentation is crucial but challenged by noisy labels in training data.
• Existing methods often fail to address pixel-dependent noise and class imbalance effectively.
• Current approaches risk discarding minority classes by using fixed thresholds for noisy label filtering.

Purpose of the Study:

• To propose a novel framework, Collaborative Learning with Curriculum Selection (CLCS), for accurate medical image segmentation.
• To address pixel-dependent noisy labels and mitigate class imbalance issues in medical imaging datasets.
• To improve data utilization by leveraging noisy samples instead of discarding them.

Main Methods:

• CLCS employs a collaborative learning framework with a two-branch network and discrepancy loss for pixel-wise noise identification.
• A curriculum dynamic thresholding approach within the Curriculum Noisy Label Sample Selection (CNS) module adaptively selects clean data samples.
• The Noise Balance Loss (NBL) module utilizes a robust loss function to leverage noisy data samples, preventing performance degradation.

Main Results:

• The CLCS framework demonstrated state-of-the-art performance on two benchmark datasets with diverse segmentation noise.
• Achieved significant improvements exceeding 3% in Dice and mean Intersection over Union (mIoU) metrics.
• Effectively handled pixel-dependent noisy labels and class imbalance, outperforming existing methods.

Conclusions:

• CLCS offers a robust solution for medical image segmentation in the presence of pixel-dependent noisy labels and class imbalance.
• The proposed method enhances segmentation accuracy by adaptively selecting clean samples and effectively utilizing noisy data.
• This framework sets a new standard for handling noisy labels in medical image segmentation tasks.