U-Shaped Medical Image Segmentation Method Based on KAN Network

随着深度学习的持续发展，基于人工智能的计算机辅助诊断技术在医学领域也有了许多进展，尤其是在医学图像分割领域。本文针对当前医学图像分割中网络复杂度高，深度学习网络可解释性差问题，提出了一种基于KAN网络的U型医学图像分割网络，KAU-Net网络。通过提出KAN通道注意力机制与KAN空间注意力机制，分别改进了U型网络中的解码器部分与跳跃连接部分，使分割网络能够在降低模型计算复杂度，提高深度学习网络可解释性。实验表明，KAU-Net网络能有效降低计算复杂度，同时提高对目标图像的分割精度，解决了由于注意力机制的加入，分割网络中计算量大大增加的问题。With the continuous development of deep learning, computer-aided diagnosis technology based on artificial intelligence has also made many advancements in the medical field, especially in the area of medical image. This paper addresses the issues of high network complexity and poor interpretability of deep learning networks in current medical image segmentation, and proposes a U-shaped medical image segmentation network based on KAN network, KAU-Net. By introducing KAN channel attention mechanism and KAN spatial attention mechanism, the decoder part and skip connection part of the U—are respectively improved, enabling the segmentation network to reduce the computational complexity of the model and enhance the interpretability of deep learning networks. Experiments show that the KAU-Net network can effectively reduce computational complexity while improving the segmentation accuracy of target images, solving the problem of significantly increased computational load due to the addition of attention mechanisms in the segmentation network.