TransUNet with unified focal loss for class-imbalanced semantic segmentation

Class imbalanceness, i.e., the inequality of the number of samples between categories, adversely affects machine learning models, including deep neural networks. In semantic segmentation, extracting a small area of minor categories with respect to the entire image includes the same problem as class imbalanceness. Such difficulties exist in various applications of semantic segmentation, including medical images. This paper proposes a semantic segmentation method that considers global features and appropriately detects small categories. The proposed method adopts TransUNet architecture and Unified Focal Loss (UFL) function; the former allows considering global image features, and the latter mitigates the harmful effects of class imbalanceness. Experimental results with real-world applications showed that the proposed method successfully extracts small regions of minor classes without increasing false positives of other classes.