DeMerge Transformer: A Learnable Decoder for Near-infrared Blurred Vessel Segmentation with Domain Adaptation

The high-fidelity identification of blood vessels plays a vital role in disease diagnosis and surgical planning. Notably, near-infrared (NIR) transillumination imaging is a safe and effective method for visualizing blood vessels. However, such images contain severe blurring that leads to difficulties obtaining sufficient well-labeled data. Inspired by the Domain Adaptation (DA) approach, this paper presents a novel decoder design called DeMerge Transformer (DMTrans) that can learn from related domains better and serve as a general upsampling approach that cooperates with multi-scale features. Our proposed model consists of a DeMerge (DM) operation and a transformer as a channel-rebuilding mechanism to complement the demerged features. DM reverses patch merging from the transformer downsampling to convert the channel information to the spatial dimension. Considering the missing channel, we further propose using a transformer architecture and a relative channel position bias to remodel the channel information in the upsampled features. The proposed technique is evaluated on a simulated blurred DRIVE dataset and a NIR vessel dataset. The results demonstrate that DMTrans outperforms state-of-the-art methods by improving most metrics using the DA method, specifically an average dice score increase of 3.47% on simulated DRIVE test data at multiple depths and 1.71% on the NIR vessel dataset.