本文为美国北卡罗来纳大学（作者：Chia-Yu Kao）的博士论文，共129页。

在过去的十年中，一种称为深度学习的机器学习技术因其能够从自然图像中提取语义而在计算机视觉中得到了广泛的应用。然而，与自然图像相比，深度学习方法在医学组织学图像分析中的效果较差。组织学图像分析包括根据细胞类型和状态对组织进行分类，其中状态之间的纹理和结构差异通常很细微。组织学图像和自然图像之间的这些定性差异使得迁移学习非常困难，并且限制了组织学图像分析中深度学习方法的使用。

本文介绍了两种新颖的深度学习体系结构，解决了这些局限性。两者都提供了帮助深度学习模型的中间提示。第一个深度学习架构是基于一个附加层（称为超层）的堆叠自动编码器构建的。超层是一个中间提示，可以捕获不同比例的图像特征。第二种结构是两层卷积神经网络（CNN），具有称为像素/区域标记的中间表示。像素/区域标记提供了一个规范化的语义描述，可以用作后续图像分类器的输入。实验表明，通过添加超层，该体系结构的性能明显优于没有中间目标的微调CNN模型。此外，实验表明，标记分类器的优点有三个方面。首先，它可以推广到其他相关的视觉任务。其次，图像分类不需要非常精确的像素标记。这种结构鲁棒，不易受噪声影响。最后，标记模型捕获低层的纹理信息并将其转换为有价值的提示。

Over the past decade, a machine learningtechnique called deep-learning has gained prominence in computer vision becauseof its ability to extract semantics from natural images. However, in contrastto the natural images, deep learning methods have been less effective foranalyzing medical histology images. Analyzing histology images involves theclassification of tissue according to cell types and states, where thedifferences in texture and structure are often subtle between states. Thesequalitative differences between histology and natural images make transferlearning difficult and limit the use of deep learning methods for histologyimage analysis. This dissertation introduces two novel deep learningarchitectures, that address these limitations. Both provide intermediate hintsto aid deep learning models. The first deep learning architecture isconstructed based on stacked autoencoders with an additional layer, called ahyperlayer. The hyperlayer is an intermediate hint that captures image featuresat different scales. The second architecture is a two-tiered ConvolutionalNeural Networks (CNN), with an intermediate representation, called apixel/region labeling. The pixel/region labels provide a normalized semanticdescription that can be used as an input to a subsequent image classifier. Theexperiments show that by adding the hyperlayer, the architecture substantiallyoutperforms fine-tuned CNN models trained without an intermediate target. Inaddition, the experiments suggest that the advantages of the labelingclassifier are threefold. First, it generalizes to other related vision tasks.Second, image classification does not require extremely accurate pixellabeling. The architecture is robust and not susceptible to the noise. Lastly,labeling model captures low-level texture information and converts them tovaluable hints.

1. 引言

2. 项目背景与相关工作

3. 使用超层跨尺度组合特征

4. 像素标记作为中间学习目标

5. 广义像素标记方法

6. 无监督像素标记层

7. 结论与未来工作展望

更多精彩文章请关注公众号：