PyMIC: A deep learning toolkit for annotation-efficient medical image segmentation (2208.09350v2)

Abstract: Background and Objective: Open-source deep learning toolkits are one of the driving forces for developing medical image segmentation models. Existing toolkits mainly focus on fully supervised segmentation and require full and accurate pixel-level annotations that are time-consuming and difficult to acquire for segmentation tasks, which makes learning from imperfect labels highly desired for reducing the annotation cost. We aim to develop a new deep learning toolkit to support annotation-efficient learning for medical image segmentation. Methods: Our proposed toolkit named PyMIC is a modular deep learning library for medical image segmentation tasks. In addition to basic components that support development of high-performance models for fully supervised segmentation, it contains several advanced components tailored for learning from imperfect annotations, such as loading annotated and unannounced images, loss functions for unannotated, partially or inaccurately annotated images, and training procedures for co-learning between multiple networks, etc. PyMIC supports development of semi-supervised, weakly supervised and noise-robust learning methods for medical image segmentation. Results: We present several illustrative medical image segmentation tasks based on PyMIC: (1) Achieving competitive performance on fully supervised learning; (2) Semi-supervised cardiac structure segmentation with only 10% training images annotated; (3) Weakly supervised segmentation using scribble annotations; and (4) Learning from noisy labels for chest radiograph segmentation. Conclusions: The PyMIC toolkit is easy to use and facilitates efficient development of medical image segmentation models with imperfect annotations. It is modular and flexible, which enables researchers to develop high-performance models with low annotation cost. The source code is available at: https://github.com/HiLab-git/PyMIC.

• The paper introduces PyMIC, a modular PyTorch-based toolkit that enables annotation-efficient segmentation using supervised, semi-supervised, and weakly supervised strategies.
• It employs advanced techniques like Uncertainty-Aware Mean Teacher and Cross-Pseudo Supervision to improve segmentation performance, notably enhancing Dice scores in medical imaging tasks.
• The toolkit reduces annotation costs and facilitates robust model development in challenges such as cardiac MRI and CATARACTS, supporting versatile segmentation tasks.

Overview of PyMIC: A Toolkit for Annotation-Efficient Medical Image Segmentation

The paper "PyMIC: A deep learning toolkit for annotation-efficient medical image segmentation" introduces PyMIC, a comprehensive modular deep learning library specifically designed for medical image segmentation. This work adds significant value to the domain by facilitating the development of models that require minimal annotation effort, addressing the challenges associated with obtaining extensive ground truth data in medical imaging.

Methodology

PyMIC is built upon the PyTorch framework, providing both fundamental and advanced components to enable efficient model development across various segmentation tasks with incomplete or imperfect annotations. The library supports:

• Fully Supervised Learning: Utilizing standard neural networks such as UNet and its variants for supervised segmentation tasks with available full annotations.
• Semi-Supervised Learning (SSL): This allows models to learn from a small subset of annotated images and a larger unannotated set, significantly reducing the need for exhaustive label annotations. Notable SSL methods include Uncertainty-Aware Mean Teacher and Cross-Pseudo Supervision.
• Weakly Supervised Learning (WSL): Enables training with sparse annotations like scribbles, incorporating specialized loss functions and regularization techniques such as entropy minimization and gated CRFs.
• Noisy Label Learning (NLL): Facilitates learning robust models from noisy annotations, using noise-robust loss functions and advanced techniques like Co-Teaching and TriNet.

Results

The PyMIC framework showcases its versatility and efficiency through its application to different segmentation tasks. For fully supervised tasks, PyMIC demonstrated competitive performance in challenges like MyoPS 2020 and CATARACTS 2020. In semi-supervised scenarios, PyMIC outperformed baseline methods with significant improvements in Dice scores in tasks such as cardiac MRI segmentation with limited annotations. Weakly supervised experiments using scribble annotations also showed marked improvements, highlighting PyMIC's potential to function with minimal supervision.

Implications

The introduction of PyMIC is a significant contribution to annotation-efficient learning in medical imaging. By providing a flexible, modular approach, it enables researchers to implement state-of-the-art models with reduced annotation costs. The comprehensive range of functionalities and ease of integration into existing pipelines make PyMIC a valuable tool in medical image computing research.

Future Directions

Continued development of PyMIC could include the integration of additional state-of-the-art network architectures and further expansion into other types of annotation-efficient learning methods like self-supervised learning and domain adaptation. Enhancements in the user interface, configurability, and interpretability could also make PyMIC an indispensable asset for the medical image analysis community.

PyMIC exemplifies a powerful, adaptable solution addressing one of the critical barriers in deploying deep learning in medical image segmentation—annotation cost. By fostering innovation and development with limited data, it aligns with the urgent need for efficient and scalable solutions in medical imaging research.