Deep Clustering for Image Data: Methods and Applications

聚类是图像处理和机器学习领域的常见任务。这种无监督学习技术调查数据点的相似性，并研究相似的数据点如何自然地分组在一起。文献中对传统环境下的聚类问题进行了广泛的研究。然而，聚类噪声、高维数据点仍然是一个开放的问题，因为许多标准算法的假设可能在高维设置中不成立。尽管深度学习已经渗透到机器学习的几乎所有方面，但关于深度聚类的出版物相对有限。基于深度学习聚类的有限研究激励着我们进行进一步的研究。目前的研究计划将通过建立理论基础和开发特定模型和特定应用的新算法，专注于图像数据分析的深度聚类。具体来说，该计划将追求以下主要技术目标：1)研究联合深度学习和聚类：除了堆叠式自编码器，由于其他深度学习模型在深度聚类中基本上未被探索，我们将研究其他流行的深度学习模型（例如卷积神经网络（CNN），深度玻尔兹曼机（DBM）和发展中的深度字典学习（DDL））的聚类性能。我们将为联合深度学习的嵌入和聚类开发端到端学习框架。我们将研究不同的深度学习模型、聚类方法、目标函数和训练/优化策略。我们还将讨论深度聚类中的模型可解释性问题。2)研究医学图像分析应用：医学图像聚类在医学成像中有很多应用。基于申请人目前的研究重点，为了测试所提出的深度聚类方法并研究其对现实医学问题的适用性，我们将重点研究医学图像分割和医学图像检索的应用。通过将本工作中开发的算法应用于医学成像中的现实问题，我们将强调它们的实际用途并了解它们的局限性。我们还需要在实践中考虑到额外的限制/挑战，开发特定于应用程序的算法。本研究的意义在于对高维数据深度聚类的模型和实践进行了研究。本研究项目的成果将对大图像数据分析的聚类方法和应用做出重大贡献。作为关键的创新加速器，大数据分析正在给许多研究和行业领域带来革命性的变化。拟议中的研究将有助于将这一愿景向前推进一小步。该研究项目将为研究生提供在相关前沿技术方面进行培训的机会。

Clustering is a common task in image processing and machine learning fields. This unsupervised learning technique investigates similarities in data points and studies how similar data points are naturally grouped together. Clustering problems have been extensively studied in the literature for traditional settings. However, clustering noisy, high-dimensional data points remains an open problem because many assumptions of standard algorithms may not hold true in the high-dimensional setting. Although deep learning has pervaded almost all aspects of machine learning, there have been relatively limited publications on deep clustering. The limited research on deep learning based clustering motivates us to carry out further research. The current research program will focus on deep clustering for image data analysis by both establishing theoretical foundations and developing model-specific and application-specific novel algorithms. Specifically, the proposed program will pursue the following main technical objectives: 1) Investigating joint deep learning and clustering: Apart from the stacked autoencoder, since other deep learning models are largely unexplored in deep clustering, we will investigate clustering performance with other popular deep learning models (e.g., convolutional neural networks (CNN), deep Boltzmann machines (DBM) and the developing deep dictionary learning (DDL)). We will develop end-to-end learning frameworks for joint deep learning-based embedding and clustering. We will investigate different deep learning models, clustering approaches, objective functions and training/optimization strategies. We will also address the model interpretability issue in deep clustering. 2) Investigating medical image analysis applications: Clustering of medical images can have many applications in medical imaging. Based on the applicant's current research focus, to test the proposed deep clustering methods and investigate their applicability to real-world medical problems, we will particularly investigate medical image segmentation and medical medical image retrieval applications. By applying the algorithms developed in this work on real-world problems in medical imaging, we will highlight their practical utility and understand their limitations. We will also need to develop application-specific algorithms given additional constraints/challenges in practice. The significance of this research lies in its focus on both the model and practice of deep clustering of high dimensional data. The outcome of this research program will make significant contributions to both the clustering methods and applications of big image data analytics. As one key innovation accelerator, big data analytics is revolutionizing many research and industry areas. The proposed research will help take this vision one tiny step further. The research program will provide an opportunity for the graduate students to be trained in related cutting-edge technologies.