Bridging the Gap Between CT and ML
We're planning to build VoxelVault, a self-hostable platform for storing, normalizing, annotating, and augmenting your CT data. It will also feature seamless integration with modern ML tools, allow for model training via GUI for non-experts, and enable automations based on incoming data and model inference results.
Stay Updated
We are currently talking to people to find out what to focus on first.
The Idea
TL;DR
Current CT software is built for human analysis first, automation second. In the limit, NDT will be fully automated. We're planning to build the software stack that helps you get there.
Problems
1) Working with large quantities of CT data for ML applications is a time sink:
- Data Collection: Viewing and organizing large quantities of scans is cumbersome as many existing tools focus on individual scan analysis.
- Annotation: Annotating 3D data is time consuming, tools that make it easier mostly focus on medical data formats.
- Data Formats: Inconsistent data formats across different CT scanners and software systems have to be normalized.
- Data Security: NDT data can be highly sensitive, so cloud-based solutions are off the table.
2) The CT and ML worlds are disconnected:
- Expertise: CT specialists lack ML expertise, and ML specialists lack CT expertise.
- Tool Integration: Traditional CT software does not integrate well with standard ML tooling.
Our Vision
1) VoxelVault, a data platform to manage CT data for ML applications:
- Format Standardization: Support for all major CT devices and reconstruction software, while trying to establish a standard format.
- Data Screening: Tools to view, filter, and combine CT scans to datasets effectively.
- Data Annotation: Labeling tools designed to make annotating 3D CT data faster.
- Data Augmentation: Extend datasets with augmented samples using a few mouse clicks.
- Self-Hostable: Choose between comfortable cloud access or secure self-hosting for maximum data control.
- Secure Collaboration: Features for access control and managing associated legal information.
2) Seamless integration with ML tooling:
- Accessible Training: Train models for common tasks like segmentation right from the GUI or using a high-level Python SDK, without having to be a ML expert.
- PyTorch Integration: A low-level Python SDK that integrates with standard machine learning tooling like PyTorch.
- Easy Deployment: Deploy models to the cloud or a local GPU cluster.
- Inference Automations: Automatically analyze new data using trained models, and trigger workflows based on results.
The Team
Get in Touch
We're currently evaluating and shaping our idea by talking to as many people as we can. So if our project sounds interesting to you, we are always happy to talk! Feel free to send us an email or book a call.