Unraveling Systematic Biases in Brain Segmentation: Insights from Synthetic Training

Romain Valabregue, Ines Khemir, Guillaume Auzias, François Rousseau, Mehdi OUNISSI

Keywords: deep neural network, Synthetic data, brain segmentation, MRI, Validation, Ground Truth, bias

Abstract:

This study examines how the quality of ground truth labels affects brain MRI segmentation models. We investigate the potential of synthetic learning to mitigate systematic biases present in training labels. Through a validation on high-quality datasets, in the Putamen region, known for systematic segmentation errors like the inclusion of parts of the Claustrum, we demonstrate the effectiveness of the synthetic data approach in correcting these errors and enhancing segmentation accuracy.
Our findings highlight the limitations of pseudo-ground truth labels derived from automated techniques and underscores the importance of precise, expert-validated labels for accurate, unbiased validation.

Submission Number: 91