U-shape-based architecture with Adjusted Bilateral Guided Aggregation Layer for nuclei image segmentation

Abstract

Medical image segmentation with AI models has been showing its incredible improvements recently and one of the most significant building blocks is U-shape architecture. With that being said, a big issue of encoder-decoder based models is the lack of semantic information during the decoding process despite the backing up of the skip connection layer. In this paper, we address this challenge with our proposed model. Inheriting the idea of spatial path and semantic path from BiSeNetV2, we replaced the skip connection between encoder and decoder blocks with Adjusted Bilateral Guided Aggregation (ABGA) layer. In addition, we also leveraged the efficiency and elegance from EffecientNet as the model’s encoder. Besides that, a new loss function based on Kulczycki 2 coefficient is introduced.