MIP Candy: A Modular PyTorch Framework for Medical Image Processing

Researchers Tianhao Fu and Yucheng Chen introduced MIP Candy (MIPCandy), a new modular PyTorch framework specifically designed for medical image processing, in a paper submitted to arXiv on February 24, 2026. The framework addresses critical challenges in medical imaging by providing a flexible solution that handles high-dimensional volumetric data, heterogeneous file formats, and domain-specific training procedures more effectively than existing tools. MIP Candy offers researchers a complete, modular pipeline spanning data loading, training, inference, and evaluation, allowing them to establish a fully functional workflow by implementing just a single method while maintaining fine-grained control over each component. The framework's central innovation is its deferred configuration mechanism, which enables runtime substitution of convolution, normalization, and activation modules without requiring subclassing. This flexibility is complemented by several built-in features including k-fold cross-validation, dataset inspection with automatic region-of-interest detection, deep supervision capabilities, exponential moving average, and multi-frontend experiment tracking integration with platforms like Weights & Biases, Notion, and MLflow. Additionally, MIP Candy provides training state recovery functionality and validation score prediction through quotient regression. An extensible bundle ecosystem further enhances the framework's utility by offering pre-built model implementations that follow a consistent trainer-predictor pattern and integrate seamlessly with the core framework without requiring modifications. The open-source project, released under the Apache-2.0 license, requires Python 3.12 or later and makes its source code and documentation publicly available through a provided URL. This comprehensive approach positions MIP Candy as a significant advancement in medical image processing software, addressing the limitations of both overly simplistic frameworks and rigid, monolithic systems.