College of Automation, Chongqing University of Posts and Telecommunications, Chongqing, 400065, China.
Emergency Department, The Second Affiliated Hospital of Chongqing Medical University, Chongqing, 400013, China; Department of Cardiovascular Surgery, Chongqing General Hospital, University of Chinese Academy of Sciences, Chongqing, 400013, China.
Comput Biol Med. 2024 Dec;183:109259. doi: 10.1016/j.compbiomed.2024.109259. Epub 2024 Oct 28.
Medical image segmentation is currently of a priori guiding significance in medical research and clinical diagnosis. In recent years, neural network-based methods have improved in terms of segmentation accuracy and become the mainstream in the field of medical image segmentation. However, the large number of parameters and computations of prevailing methods currently pose big challenges when employed on mobile devices. While, the lightweight model has great potential to be ported to low-resource hardware devices for its high accuracy. To address the above issues, this paper proposes a lightweight medical image segmentation method combining Transformer and Multi-Layer Perceptron (MLP), aiming to achieve accurate segmentation with lower computational cost. The method consists of a multi-scale branches aggregate module (MBA), a lightweight shift MLP module (LSM) and a feature information share module (FIS). The above three modules are integrated into a U-shaped network. The MBA module learns image features accurately by multi-scale aggregation of global spatial and local detail features. The LSM module introduces shift operations to capture the associations between pixels in different locations in the image. The FIS module interactively fuses multi-stage feature maps acting in skip connections to make the fusion effect finer. The method is validated on ISIC 2018 and 2018 DSB datasets. Experimental results demonstrate that the method outperforms many state-of-the-art lightweight segmentation methods and achieves a balance between segmentation accuracy and computational cost.
医学图像分割在医学研究和临床诊断中具有先验的指导意义。近年来，基于神经网络的方法在分割精度方面取得了进步，成为医学图像分割领域的主流。然而，当前主流方法的大量参数和计算量在移动设备上应用时带来了很大的挑战。而轻量化模型具有很大的潜力，可以针对低资源硬件设备进行移植，因为它具有很高的准确性。针对上述问题，本文提出了一种结合 Transformer 和多层感知机（MLP）的轻量化医学图像分割方法，旨在以较低的计算成本实现精确分割。该方法由多尺度分支聚合模块（MBA）、轻量级移位 MLP 模块（LSM）和特征信息共享模块（FIS）组成。上述三个模块集成到一个 U 形网络中。MBA 模块通过全局空间和局部细节特征的多尺度聚合，准确地学习图像特征。LSM 模块引入移位操作来捕捉图像中不同位置像素之间的关联。FIS 模块通过交互融合在跳跃连接中作用的多阶段特征图，使融合效果更加精细。该方法在 ISIC 2018 和 2018 DSB 数据集上进行了验证。实验结果表明，该方法优于许多最新的轻量化分割方法，在分割精度和计算成本之间取得了平衡。
