Strategies for tackling the class imbalance problem of oropharyngeal primary tumor segmentation on magnetic resonance imaging

Abstract

No significant differences were observed when training with different loss functions. The two-stage approach outperformed the end-to-end 3D U-Net.