基于深度学习的癌变组织显微高光谱图像分类  

Classification of Microscopic Hyperspectral Images of Cancerous Tissue Based on Deep Learning

作　　者：张勇 黄丹飞[1,2] 张乐超[1,2] 张丽丽 周尧 唐鸿宇[1,2] Zhang Yong;Huang Danfei;Zhang Lechao;Zhang Lili;Zhou Yao;Tang Hongyu(School of Optoelectronic Engineering,Changchun University of Science and Technology,Changchun 130022,Jilin,China;Zhongshan Institute,Changchun University of Science and Technology,Zhongshan 528400,Guangdong,China)

机构地区：[1]长春理工大学光电工程学院,吉林长春130022 [2]长春理工大学中山研究院,广东中山528400

出　　处：《激光与光电子学进展》2024年第18期237-244,共8页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61893096014)。

摘　　要：在因式分解卷积神经网络和残差结构的思想下,利用膨胀卷积,并添加注意力机制,提出了一种融合混合注意力机制模块的残差分解卷积神经网络(CBAM-RFNet)。该网络主要是把传统3×3的二维卷积因式分解为3×1和1×3的两个一维卷积串联,不仅增加了网络模型的深度,还减少了参数,是一种轻量级的网络模型。在显微高光谱成像系统采集的甲状腺癌显微高光谱图像上的实验结果显示,与其他深层的神经网络相比,提出的网络能有效提升显微高光谱图像的分类精度,其分类的总体准确率为98.23%,F1值为98.66%,Kappa系数为0.909。Based on the idea of factorization neural network and residual structure,a convolutional block attention module for residual factorized of convolutional neural networks(CBAM-RFNet)is proposed by expansive convolution and adding attention mechanism.In this network,the traditional 3×3 two-dimensional convolution is decomposed into two onedimensional convolution of 3×1 and 1×3 and connect them in series,which not only increases the depth of the network model,but also reduces the parameters,the network is a lightweight network model.The experimental results on thyroid cancer images collected by microhyperspectral imaging system show that,compared with other deep neural networks,the proposed network can effectively improve the classification accuracy of microhyperspectral images,with the overall accuracy of 98.23%,F1 value of 98.66%,and Kappa coefficient of 0.909.

关 键 词：显微高光谱图像分类 因式分解卷积 膨胀卷积 注意力机制 轻量级网络 

分 类 号：R736.1[医药卫生—肿瘤] TP391.9[医药卫生—临床医学] TP183[自动化与计算机技术—计算机应用技术]