A Lightweight CNN-Transformer Network with Agent Attention Mechanism for 3D Whole Heart Segmentation

In recent years, heart diseases have become one of the major global health issues, emphasizing the importance of 3D whole heart segmentation technology for accurate and efficient treatment. Currently, 3D whole heart segmentation algorithms often combine Transformer technologies with 3D Convolutional Neural Networks (3D CNNs), leveraging the self-attention mechanism of Transformers for global information modeling and 3D CNNs for capturing local contextual information. However, the high computational complexity of Transformers leads to huge computational resource wastage. To address this issue, this work proposes a novel whole heart segmentation algorithm named AgentTransBTS, based on an improved 3D Agent Attention mechanism and the TransBTS framework. This algorithm utilizes TransBTS as the baseline model, capitalizing on its deep feature extraction and information integration capabilities to capture global dependencies in images. To reduce computational complexity without sacrificing the model's expressive power, an enhanced 3D Agent Attention mechanism is introduced. Comprehensive experiments were conducted on the Multi-Modality Whole Heart Segmentation (MM-WHS) 2017 dataset to evaluate the proposed algorithm. The experimental results demonstrate that our proposed method achieved an average Dice coefficient of 91.9% on the CT dataset and 79.9% on the MRI dataset of the MM-WHS 2017 challenge. Furthermore, our algorithm's FLOPs are 217G, representing an approximate 9.2% improvement in computational efficiency compared to the baseline network.