CombTR is an ensemble model developed to automatically segment 13 organs based off of 3D CT scans. The following models are used:
- UNETR
- SwinUNETR
- SegResNet
CombTR uses a "stacking" model architecture, where each model's output is the input into a meta-learner, which learns from the ensemble's mistakes and corrects them over time. This model beat previous research from NVIDIA and Vanderbilt University by 5% in DICE score.