A Stomach Recognition Method Based on TransUnet and Dual Attention Mechanism

To address the issue of insufficient local detail and incomplete global information capture in stomach organ recognition on abdominal CT images, a dual attention mechanism that fuses channel and spatial attention was proposed and embedded within a Convolutional Neural Network (CNN)-Transformer hybrid architecture. In the encoder, a channel attention mechanism is introduced that dynamically adjusts the weights of each channel to enhance the model's feature extraction capability. Meanwhile, a spatial attention mechanism is embedded within the skip connections to enable the model to focus on key regions and better preserve spatial details during resolution recovery. The publicly available Synapse dataset was used for validation, and comparative experiments were performed with state-of-the-art models—including Unet, Unet++, Medical Transformer, and TransUnet—and the results indicated that the proposed method not only achieved a competitive Dice coefficient but also demonstrated significant enhancements in boundary accuracy, improving the Dice coefficient by 3.58% and the Hausdorff distance by 29.5% relative to the baseline model.