基于Swin Transformer的疟疾细胞图像识别研究-《中国医学物理学杂志》

文章信息/Info

作者:: 贺鹏飞1; 马建飞1; 李成林1; 张桐敬2; 粱大伟3; 1.烟台大学物理与电子信息学院，山东烟台 264005； 2.华能山东烟台发电有限公司烟台发电厂，山东烟台 264002； 3.烟台市食品药品检测中心，山东烟台 264000

Author(s):: HE Pengfei1; MA Jianfei1; LI Chenglin1; ZHANG Tongjing2; LIANG Dawei3; 1. School of Physics and Electronic Information, Yantai University, Yantai 264005, China 2. Yantai Power Plant, Huaneng Shandong Power Generation Co., Ltd, Yantai 264002, China 3. Yantai Center for Food and Drug Control, Yantai 264000, China

Keywords:: Keywords: medical image classification Swin Transformer pseudo-color image processing residual structure SW-MSA

摘要:: 为了协助医务人员更准确、更快速地诊断疟疾，提出一种基于Swin Transformer（SwinT）的疟疾细胞图像识别方案。方案采用伪彩色图像增强算法对血片图像进行预处理，以突出图像的颜色对比度，并引入SwinT模型作为主干网络，解决下采样固定和全局信息无法交互的问题，同时引入卷积层对图像进行线性变换，构建残差网络解决梯度消失和梯度爆炸问题。实验表明，与图像量化等其他图像增强方法相比，本文方法增强了疟疾细胞图像的色彩对比度，改进后方案的准确率达到99.7%，高于现有文献方法，可以对疟疾的辅助治疗带来更有价值的支持。

Abstract:: Abstract: A Swin Transformer (SwinT)-based scheme for malaria cell image recognition is proposed to assist medical personnel in diagnosing malaria more accurately and quickly. The scheme pre-processes blood films with a pseudo-color image enhancement algorithm to highlight the color contrast, and uses SwinT model as the backbone network to solve the problems of fixed downsampling and the inability to interact with global information, while introducing a convolutional layer for linear transformation and constructing a residual network to address the issues of gradient disappearance and gradient explosion. Experiments show that compared with other image enhancement methods such as image quantization, the proposed method enhances the color contrast of malaria cell images. The accuracy of the improved scheme reaches 99.7%, higher than the existing literature methods and bringing more valuable support to the adjunctive treatment of malaria.