ECA-TFUnet: A U-shaped CNN-Transformer network with efficient channel attention for organ segmentation in anatomical sectional images of canines

Abstract

Automated organ segmentation in anatomical sectional images of canines is crucial for clinical applications and the study of sectional anatomy. The manual delineation of organ boundaries by experts is a time-consuming and laborious task. However, semi-automatic segmentation methods have shown low segmentation accuracy. Deep learning-based CNN models lack the ability to establish long-range dependencies, leading to limited segmentation performance. Although Transformer-based models excel at establishing long-range dependencies, they face a limitation in capturing local detail information. To address these challenges, we propose a novel ECA-TFUnet model for organ segmentation in anatomical sectional images of canines. ECA-TFUnet model is a U-shaped CNN-Transformer network with Efficient Channel Attention, which fully combines the strengths of the Unet network and Transformer block. Specifically, The U-Net network is excellent at capturing detailed local information. The Transformer block is equipped in the first skip connection layer of the Unet network to effectively learn the global dependencies of different regions, which improves the representation ability of the model. Additionally, the Efficient Channel Attention Block is introduced to the Unet network to focus on more important channel information, further improving the robustness of the model. Furthermore, the mixed loss strategy is incorporated to alleviate the problem of class imbalance. Experimental results showed that the ECA-TFUnet model yielded 92.63% IoU, outperforming 11 state-of-the-art methods. To comprehensively evaluate the model performance, we also conducted experiments on a public dataset, which achieved 87.93% IoU, still superior to 11 state-of-the-art methods. Finally, we explored the use of a transfer learning strategy to provide good initialization parameters for the ECA-TFUnet model. We demonstrated that the ECA-TFUnet model exhibits superior segmentation performance on anatomical sectional images of canines, which has the potential for application in medical clinical diagnosis.

Keywords

• anatomical sectional images of canines
• segmentation
• transformer
• efficient channel attention
• unet network
• transfer learning