LWV-UNet: Lightweight Vision Transformer UNet for Nasopharyngeal Carcinoma Image Segmentation

Nasopharyngeal carcinoma (NPC) is a common malignancy of the head and neck. Deep learning-based image segmentation of NPC can facilitate clinical diagnosis. However, these models are often characterized by a large number of parameters and high computational complexity. To address this issue, we propose a Lightweight Vision Transformer UNet (LWV-UNet) model. This model combines the strengths of convolutional neural networks and Transformers, and achieves lightweight characteristics by using Inverted Residual (IR) blocks and MobileViTv2 blocks to replace the encoder. The MobileViTv2 block enables global feature modeling, allowing it to capture complex image structures. Additionally, we use IR blocks in the skip connections to learn information from different feature maps. Finally, the decoder, composed of IR blocks, generates the predicted segmentation map. For the loss function, we apply deep supervision to enhance the network's ability to learn feature information. We evaluated the proposed method on an in-house NPC dataset. Experimental results demonstrate that the proposed LWV-UNet achieves comparable segmentation performance to other methods while reducing computational cost and the number of parameters.