PyMIC: A deep learning toolkit for annotation-efficient medical image segmentation.

Area of Science:

• Medical image analysis
• Deep learning in healthcare
• Computational pathology

Background:

• Medical image segmentation is crucial for computer-assisted diagnosis and treatment.
• Current deep learning toolkits primarily rely on fully supervised methods requiring extensive pixel-level annotations.
• Acquiring accurate annotations is time-consuming and costly, necessitating annotation-efficient learning strategies.

Purpose of the Study:

• To develop an open-source deep learning toolkit, PyMIC, supporting annotation-efficient learning for medical image segmentation.
• To accelerate and simplify the development of deep learning models with limited annotation budgets.
• To enable learning from partial, sparse, or noisy annotations.

Main Methods:

• PyMIC is a modular deep learning library built on the PyTorch framework.
• It includes components for fully supervised segmentation and advanced features for learning from imperfect annotations.
• Supports semi-supervised, weakly supervised, and noise-robust learning methods.

Main Results:

• Demonstrated competitive performance in fully supervised learning.
• Achieved semi-supervised cardiac structure segmentation using only 10% annotated images.
• Showcased weakly supervised segmentation with scribble annotations and noise-robust learning for chest radiographs.

Conclusions:

• PyMIC facilitates efficient development of medical image segmentation models with imperfect annotations.
• The toolkit is modular and flexible, enabling high-performance model development at a low annotation cost.
• Source code is publicly available for research use.