ThresholdNet

ThresholdNet enhances automatic segmentation of polyps in endoscopy images to support improved early detection and treatment of colorectal cancer.

Key Features:

• Confidence-Guided Manifold Mixup (CGMMix): Performs manifold mixup at both image and feature levels to augment training data and increase sample diversity.
• Confidence-guided decision boundary adjustment: Uses confidence information to adaptively shift the decision boundary away from under-represented polyp classes to mitigate class imbalance.
• Mixup Feature Map Consistency (MFMC) loss: Enforces consistency of feature maps between original and mixed samples during training.
• Mixup Confidence Map Consistency (MCMC) loss: Enforces consistency of confidence maps between original and mixed samples during training.
• Two-branch segmentation and threshold learning: Integrates a segmentation branch and a threshold-learning branch with an alternative training strategy.
• Threshold Map Supervision Generator (TMSG): Provides supervision for the learned threshold map to optimize the threshold branch.
• Threshold calibration: Calibrates segmentation outputs using the learned threshold map instead of fixed likelihood thresholds (e.g., 0.5).

Scientific Applications:

• Polyp segmentation in endoscopy images: Applied to polyp segmentation tasks for colorectal cancer screening using endoscopic imagery.
• Benchmark evaluation: Demonstrated on the EndoScene dataset (Dice 87.307%) and the WCE polyp dataset (Dice 87.879%).
• Handling limited annotations and class imbalance: Addresses challenges arising from limited annotated datasets and imbalanced polyp classes via CGMMix and confidence-guided learning.

Methodology:

Manifold mixup at image and feature levels (CGMMix) with confidence guidance to adjust decision boundaries; MFMC and MCMC consistency losses applied to mixed data; a two-branch alternative training strategy with a Threshold Map Supervision Generator (TMSG) to learn and apply a threshold map for segmentation calibration.