本文为奥地利维也纳技术大学（作者：Philipp Seeböck）的硕士论文，共161页。

机器学习应用于医学成像领域，包括计算机辅助诊断、图像分割、图像配准、图像融合、图像引导治疗、图像注释和图像数据库检索。深度学习方法是机器学习中的一套算法，它试图自动学习多个层次的表示和抽象，帮助理解复杂的数据。这反过来又导致了理解和检查深层学习方法特征的必要性，以便能够以适当的方式应用和改进这些方法。

本研究的目的是评估医学领域的深度学习方法，并了解深度学习方法（随机递归支持向量机、堆积稀疏自动编码器、堆积去噪自动编码器、K-均值深度学习算法）是否优于其他最先进的方法（K-最近邻、支持向量机）。在两个分类任务中，所有方法在手写数字（MINIST）和医学（PULMO）数据集上进行了评估。除了准确度和运行时间方面的评估外，本文还对所学习的特性进行了定性分析，并对评估方法提出了实际建议。这将有助于评估方法的应用和改进。结果表明，在两个数据集的所有评估方法中，堆积稀疏自动编码器、堆积去噪自动编码器和支持向量机的精度最高。如果可用的计算资源允许使用这些方法，则是可取的。相比之下，随机递归支持向量机在两个数据集上的训练时间最短，但比上述方法获得的精度较差。这意味着，如果计算资源有限，并且运行时间是一个重要问题，那么应该使用随机递归支持向量机。

Machine learning is used in the medicalimaging field, including computer-aided diagnosis, image segmentation, imageregistration, image fusion, image-guided therapy, image annotation, and imagedatabase retrieval. Deep learning methods are a set of algorithms in machinelearning, which try to automatically learn multiple levels of representationand abstraction that help make sense of data. This in turn leads to thenecessity of understanding and examining the characteristics of deep learningapproaches, in order to be able to apply and refine the methods in a properway. The aim of this work is to evaluate deep learning methods in the medicaldomain and to understand if deep learning methods (random recursive supportvector machines, stacked sparse auto-encoders, stacked denoising auto-encoders,K-means deep learning algorithm) outperform other state of the art approaches(K-nearest neighbor, support vector machines, extremely randomized trees) ontwo classification tasks, where the methods are evaluated on a handwrittendigit (MNIST) and on a medical (PULMO) dataset. Beside an evaluation in termsof accuracy and runtime, a qualitative analysis of the learned features andpractical recommendations for the evaluated methods are provided within thiswork. This should help improve the application and refinement of the evaluatedmethods in future. Results indicate that the stacked sparse auto-encoder, thestacked denoising auto-encoder and the support vector machine achieve thehighest accuracy among all evaluated approaches on both datasets. These methodsare preferable, if the available computational resources allow to use them. Incontrast, the random recursive support vector machines exhibit the shortesttraining time on both datasets, but achieve a poorer accuracy than the aforementioned approaches. This implies that if the computational resources arelimited and the runtime is an important issue, the random recursive supportvector machines should be used. 

1 引言

2 机器学习

3 深度学习

4 算法实现

5 基于MNIST数据集的学习方法比较

6 基于PULMO数据集的学习方法比较

7 总结与讨论

8 结论与未来工作展望

附录A MNIST数据集的更多结果

附录B PULMO数据集的更多结果

更多精彩文章请关注公众号：