MIPCandy: Accelerating Medical AI Research

During my internship at UTMIST (University of Toronto Machine Intelligence Student Team), Project Neura, I contribute to the development of MIPCandy, an open-source infrastructure framework designed to streamline medical image processing workflows.

The Challenge

Medical imaging research often requires significant engineering effort before researchers can focus on their core innovations. Setting up training pipelines, integrating evaluation metrics, and managing experiment tracking can take weeks of development time.

What is MIPCandy?

MIPCandy is a PyTorch-based framework that enables researchers to "build a complete experiment pipeline for your medical imaging model in 10 seconds." The framework provides out-of-the-box integration of neural network architectures with their corresponding training, inference, and evaluation pipelines.

Key Features

• Pre-integrated Architectures: Support for state-of-the-art medical imaging models including UNet, UNetR, UNet++, and MedNeXt
• Flexible Training System: Configurable presets for different medical imaging tasks
• Advanced Data Processing: Comprehensive data preprocessing and augmentation pipelines
• Random Patch Support: Efficient handling of large medical images through intelligent patch sampling
• Dashboard Integration: Seamless integration with Notion, WandB, and TensorBoard for experiment tracking

My Contributions

As a research intern, I focus on several critical components of the framework:

• Developing model adapters for various segmentation architectures (UNet, UNetR, UNet++, MedNeXt)
• Implementing training presets for different medical imaging modalities
• Building robust data processing pipelines
• Adding random patch sampling support for efficient training on high-resolution medical images

Impact

MIPCandy supports multiple medical imaging tasks including classification, segmentation, and detection across various modalities such as MRI and CT scans. By reducing the engineering overhead, the framework allows researchers to focus on advancing medical AI rather than rebuilding infrastructure.

This project represents an important step in democratizing medical AI research, making it more accessible to researchers and accelerating the development of diagnostic tools that can improve patient outcomes.