Consistent Brain MRI Scans Across Devices, No Model Retraining
Unpaired image-to-image translation to standardize brain MRI scans from multiple hospitals
About the customer
Roche, a global pharmaceutical leader known for its work in oncology and immunology, needed to standardize brain MRI images coming from different hospitals and scanners. Consistent images matter because Roche’s pre-trained models for brain tumor classification and segmentation (outlining the tumor in the scan) only work well on data that looks like what they were trained on. The goal was to make MRI images from different machines consistent, so these models keep performing.
What needed solving
Inconsistent scans
MRI images varied across multiple hospitals and imaging devices
Model performance
These variations hurt machine learning models that had been trained on standardized data
Tumor detection
Non-uniform T1 and T2 FLAIR scans (two common MRI scan types) prevented accurate tumor detection
No paired data
Images had to be standardized without paired data, because matching pairs of the same scan from different hospitals were not available
How we built it
Model choice
After a literature review, chose CycleGAN, a deep learning model designed for unpaired image translation.
Image translation
Converts 2D slices taken from 3D brain MRI images into a standard format that Roche’s pre-trained models can use.
Training and validation
Trained on T1 and T2 FLAIR scans, tested rigorously, and validated by MRI experts.
Easy integration
Packaged with Docker so it can be deployed in Roche’s existing systems.
Roche’s diagnostic models now work across data sources without re-training
Technologies
Domains
Have a Query? Let's Have a Conversation
Do you have a project in mind? We’d love to know what it is. Email us or send a message from our contact page.
Have a similar problem?
Tell us what you are trying to solve and our team will get back to you.