摘要
胃癌是全球第五常见的恶性肿瘤并死亡率较高,严重威胁人类的生命健康。因此,早期识别胃癌病变对早期胃癌诊断至关重要。显微高光谱成像技术作为一种新兴技术,可以在微观层面同时获取生物组织丰富的光谱信息和空间信息,为早期病理切片诊断提供了一种新的思路。利用显微高光谱成像系统,采集了在400~1000 nm波段范围的胃癌显微高光谱病理图像,通过光谱校正等预处理构建了包含230张图像的胃癌显微高光谱数据集。尽管基于空间注意力的方法在图像分类、分割等领域已取得了显著成果,但在处理高光谱图像时仍面临计算复杂度高以及光谱信息利用不充分的问题。为此,提出了基于卷积和注意力机制的混合双分支Transformer(MDBT)的特征提取主干网络模型。该模型通过交替应用空间混合模块和通道混合模块,实现块间和块内的空间和通道特征聚合。具体而言,设计了窗口注意力和卷积双分支以及空间和通道交互结构。这种设计不仅降低了计算复杂度,还通过卷积交互实现了窗口间信息交互和特征融合,从而克服了窗口注意力感受野受限的问题,进一步提高了Transformer的全局建模能力。在进行图像分割实验中,采用UperNet模型作为解码头网络对主干网络提取得到的特征进行还原,以得到最终的分割结果。在采集得到的胃癌高光谱数据集上进行了五折交叉验证实验,结果表明本模型的平均mDice和mIoU分别达到85.39、74.66,性能优于目前UNet、Swin、PVT、VIT等主流图像分割网络模型。同时设计一系列消融实验,验证本文提出空间和通道双混合模块、卷积与窗口注意力双分支等结构对实验结果的优化效果。实验结果表明本文提出的MDBT模型能够有效利用高光谱图像丰富的空间和光谱信息,提高胃癌图像分割准确率,证明显微高光谱成像技术在胃癌诊断方面具有一定的研究意义和应用价值。
Gastric cancer is the third leading cause of cancer-related deaths globally,posing a serious threat to human life and health.Therefore,early identification of gastric cancer lesions is crucial for early diagnosis of gastric cancer.As an emerging technique,microscopic hyperspectral imaging technology can simultaneously obtain rich spectral information and spatial information of biological tissues at the microscopic level,providing a new approach for early pathological slice diagnosis.In this paper,gastric cancer microscopic hyperspectral images in the range of 400~1000 nm were collected using a microscopic hyperspectral imaging system.The gastric cancer microscopic hyperspectral dataset containing 230 images was constructed through preprocessing,such as spectral calibration.Although spatial attention-based methods have achieved significant results in image classification,segmentation,and other fields,they still face challenges of high computational complexity and insufficient utilization of spectral information when dealing with hyperspectral images.Therefore,this paper proposes a backbone network model based on convolution and attention mechanism called Mixing Dual-Branch Transformer(MDBT).This model achieves spatial and channel feature aggregation between blocks and within blocks by alternately applying spatial and channel mixing modules.Specifically,this paper designs window attention,convolution dual branches,and spatial and channel interaction structures.This design not only reduces computational complexity but also achieves window-to-window information interaction and feature fusion through convolutional interaction,overcoming the limitation of window attention's receptive field and further improving the global modeling ability of the Transformer.In the image segmentation experiments,we adopt the UperNet model as the decode head network to reconstruct the features extracted by the backbone network to obtain the final segmentation results.Five-fold cross-validation experiments were conducted on the collected gastric cancer hyperspectral dataset,and the results show that the average priceand mIoU of this paper's model reach 85.39 and 74.66,respectively,outperforming mainstream image segmentation network models such as UNet,Swin,PVT,and VIT.Meanwhile,ablation experiments are designed to verify the optimization effects of the proposed spatial and channel dual mixing modules,convolution,window attention dual branches,and other structures on experimental results.Experimental results demonstrate that the proposed MDBT model can effectively utilize hyperspectral images'rich spatial and spectral information,improve the accuracy of gastric cancer image segmentation,and prove the research significance and application value of microscopic hyperspectral imaging technology in gastric cancer diagnosis.
作者
张然
金伟
牟颖
于丙文
柏怡文
邵益波
平金良
宋鹏涛
何湘漪
刘飞
付琳琳
ZHANG Ran;JIN Wei;MU Ying;YU Bing-wen;BAI Yi-wen;SHAO Yi-bo;PING Jin-liang;SONG Peng-tao;HE Xiang-yi;LIU Fei;FU Lin-lin(College of Control Science and Engineering,Zhejiang University,Hangzhou 310058,China;Huzhou Institute of Zhejiang University,Huzhou 313000,China;Huzhou Central Hospital,Huzhou 313000,China)
出处
《光谱学与光谱分析》
北大核心
2025年第2期551-557,共7页
Spectroscopy and Spectral Analysis
基金
国家自然科学基金面上项目(32071481)
湖州市科技局公益性应用研究重点项目(2021GZB01)资助。
作者简介
张然,1998年生,浙江大学控制科学与工程学院硕士研究生,e-mail:22132072@zju.edu.cn;通讯作者:平金良,E-mail:pjl0173@163.com。
