【作者】 徐斌；

【导师】 纪禄平；

【作者基本信息】 电子科技大学 ， 计算机科学与技术， 2023， 硕士

【摘要】 WSI(While Slide Imaging)医学图像分类是一种新兴的数字病理学研究方向,旨在利用机器学习技术对组织样本的高分辨率数字图像进行分析。该领域的研究者们利用计算机视觉技术和深度学习模型,实现了对WSI图像中的细胞、组织、器官等结构的准确识别和定位,从而帮助医生更加快速、准确地做出病理诊断。当前,WSI医学图像分类已广泛应用于癌症、炎症、神经病理学等疾病的诊断和治疗,具有较高的临床应用前景。由于整个幻灯片图像的千兆像素大小,基于患者WSI对癌症的分类是一个具有挑战性的问题。目前,大多数现有WSI分类方法通常基于多实例学习(MIL)网络。在MIL算法中,WSI被组织成许多实例袋(bag)和标签。一个实例袋可以由多个实例组成,每个实例都有一个相应的特征向量。每个实例袋可以有多个标签,标签表明该实例袋是否属于某个特定的类别。即首先将整个图像视为一个包,图像的切片视为多个实例,在网络中进行训练预测,得出结果。在MIL网络中,仍有一些值得进一步探索的重要研究点,如WSI切片特征组织方式和不同规模的特征利用,以及不同任务间的网络学习。针对以上问题,本文在当前特征组织方式和MIL网络上做了研究与改进,探究了不同规模特征的利用以及MIL网络中注意力机制和多任务学习的应用,主要研究工作如下:(1)不同尺度的WSI特征提取和组织:探究了不同放大倍数下的WSI图像的特征提取和融合方案对于MIL网络效果的影响。WSI切片通常有不同尺度的特征,只利用一种尺度会造成信息过少。利用多层不同尺度的特征可以有效弥补不同尺度特征的缺点,将多种特征进行融合,从而得到信息更加丰富的特征。(2)引入通道注意力:探究了引入注意力机制之后对于WSI图像诊断的实际影响。传统卷积神经网络结构相对简单,可以训练的参数较少,且与特征数据相关性不高。注意力机制使得模型更关注那些对当前任务更重要的信息,从而提高模型的性能。通过引入通道注意力可以提高权重与特征数据的相关性,从而得到更精细的特征表达。(3)多任务学习模型:一个模型对单个任务训练时,会因为过拟合等原因导致结果不够理想。针对这个问题,我们提出了多任务学习模型。不同任务间使用同样的底部模型,再针对不同任务特性,采用适合任务的特定模型。这样可以使不同任务感受到任务差异和特征分布的影响,避免过拟合,提升效果。更多还原

【Abstract】 Medical image classification is a new research direction of digital pathology,which aims to analyze and classify high-resolution digital images of tissue samples using artificial intelligence and machine learning technology.Using computer vision technology and deep learning model,researchers in this field have realized the accurate identification and location of cells,tissues,organs and other structures in WSI images,thus helping doctors make more rapid and accurate pathological diagnosis.At present,WSI medical image classification has been widely used in the diagnosis and treatment of cancer,inflammation,neuropathy and other diseases,and has a high clinical application prospect.Due to the Giga-pixel size of whole slide image(WSI),the cancer classification by patient WSI has always been a challenging problem.Currently,most existing methods for WSI classification usually are based on multiple instance learning(MIL)networks.In the MIL algorithm,WSI is organized into many instance bags and tags.An instance bag can be composed of multiple instances,and each instance has a corresponding feature vector.Each instance bag can have multiple labels,which indicate whether the instance bag belongs to a specific category.That is,first of all,the whole image is regarded as a package,and the slice of the image is regarded as multiple instances,and then the training prediction is carried out in the network to obtain the results.In the MIL network,there are still some important research points worthy of further exploration,such as the organization of WSI slice features and the use of features of different scales,as well as the network learning between different tasks.In response to the above problems,this thesis has made research and improvement on the current feature organization and MIL network,and explored the use of features of different scales and the application of attention mechanism and multi-task learning in MIL network.The main research work is as follows:(1)WSI feature extraction and organization at different scales: explore the impact of feature extraction and fusion schemes of WSI images at different magnification on the effect of MIL network.WSI slices usually have features of different scales.Using only one scale will cause too little information,and the feature scales of previous research methods are too different.The use of multi-level features with different scales can effectively compensate for the shortcomings of features with different scales,and fuse multiple features to obtain more informative features.(2)Introduce channel attention: explore the actual impact on WSI image diagnosis after introducing attention mechanism.The traditional convolutional neural network has relatively simple structure,less parameters that can be trained,and low correlation with feature data.The attention mechanism makes the model pay more attention to the information that is more important to the current task,thus improving the performance of the model.By introducing channel attention,the correlation between weight and feature data can be improved,and more detailed feature expression can be obtained.(3)Multi-task learning model: when a model trains a single task,the result is not ideal due to over-fitting and other reasons.To solve this problem,we propose a multitask learning model.Use the same bottom model between different tasks,and then use the specific model suitable for different tasks according to the characteristics of different tasks.This can make different tasks feel the impact of task differences and feature distribution,avoid over-fitting and improve the effect.更多还原