UCITransNet: A Deep Neural Model with FC-CRFs and Object-Level Data Augmentation for Cell Instance Segmentation
Zequan Liu, Suhong Wang, Zidong Chen, Zhifeng Xiao · 2023
U-Net and its variants have been the mainstream model paradigm for a wide spectrum of semantic segmentation tasks. A recent proposal was to replace the skip connections of U-Net with a multi-scale Channel Cross fusion with Transformer (CCT) module, forming a so-called UCTransNet architecture with SOTA performance on several tasks. The characteristics of the multi-head attention module adopted in CCT lead to the fusion of multi-channel information, but it also loses some valuable feature information in the fusion calculation process, putting much pressure on the backbone network. We propose Channel-wise Isomerism fusion Transformer (CIT) module to replace the CCT module, yielding the UCITransNet model. We adopt an FC-CRF module as a single channel enhancement module to work in coordination with the CCT. The motivation has been retaining the function of channel integration as much as possible while taking into account the significance of conveying multi-scale features. In addition, we propose an object-level data augmentation method, named cell instance augmentation (CIA), to enhance the quantity and diversity of the training data, coupled with the enhanced model architecture. The proposed method was validated on two datasets, including the Gland Segmentation in Colon Histology Images Challenge (GlaS) dataset and the Multi-Organ Nucleus Segmentation Challenge (MoNuSeg) dataset using the Dice and IoU performance metrics. On both datasets, our method presents superior performance compared to the SOTA and other benchmark methods.