New method detects label bias in image segmentation without clean data
Researchers adapt Confident Learning to uncover hidden demographic biases in training labels.
Deep Dive
Researchers present a data-centric adaptation of Confident Learning to detect label bias in image segmentation without requiring clean, unbiased ground truth. By comparing training labels to model predictions, they isolate directional errors and measure bias where standard metrics like Dice fail. Evaluated on three datasets, the framework reliably mitigates bias and achieves equitable performance across demographic subgroups.
Key Points
- Adapts Confident Learning to image segmentation, detecting label bias without clean ground truth annotations.
- Bias measured by isolating directional errors; standard overlap metrics like Dice fail to capture them.
- Evaluated on three datasets (synthetic to real-world), achieving equitable performance across subgroups.
Why It Matters
Enables fairer medical imaging and autonomous driving models by exposing hidden label biases in training data.