PyMIC

PyMIC provides annotation-efficient deep learning methods for medical image segmentation to enable training from partial, sparse, or noisy pixel-level annotations using the PyTorch framework.

Key Features:

• PyTorch implementation: Implemented on the PyTorch framework for deep learning model development.
• Annotation-efficient learning: Supports training from partial, sparse, and noisy pixel-level annotations to reduce annotation requirements.
• Modular design: Modular components enable fully supervised, semi-supervised, weakly supervised, and noise-robust learning strategies.
• Data handling: Supports loading both annotated and unannotated images for mixed-supervision settings.
• Specialized loss functions: Includes loss functions tailored for unannotated, partially annotated, or inaccurately annotated images.
• Co-learning: Facilitates co-learning between multiple networks to improve performance under limited annotations.

Scientific Applications:

• General medical image segmentation: Enables development of segmentation models for computer-assisted diagnosis and treatment with limited annotated data.
• Fully supervised segmentation: Demonstrated competitive results in fully supervised learning scenarios.
• Semi-supervised cardiac structure segmentation: Enables cardiac structure segmentation with only 10% of training images annotated.
• Weakly supervised segmentation with scribbles: Supports segmentation using scribble annotations as weak supervision.
• Noisy-label chest radiograph segmentation: Supports learning from noisy labels for chest radiograph segmentation.

Methodology:

Implemented in PyTorch; supports loading annotated and unannotated images, employs specialized loss functions for unannotated, partially or inaccurately annotated images, and facilitates co-learning between multiple networks for fully supervised, semi-supervised, weakly supervised, and noise-robust learning.