Sparse Coding and Auto-encoders for Advanced and Robust Processing of Biomedical Images

DSP在医学图像处理中的应用已经成为当今医疗中心几乎每一个程序、机器和系统中不可或缺的一部分。提出的研究计划的总体目标是开发新的信号和图像处理算法，并为医学成像系统实现鲁棒和可靠的技术。深度卷积神经网络和压缩感知方法将在本研究中进行研究。研究的基础和理论要素包括图像分析、生成、增强、配准、重建和建模。在动态心脏电影磁共振成像（MRI）中，扫描时间较长，因为空间频率信息是顺序检索的，只有将所有频率信息收集到一起才能形成图像。因此，在这些获取的图像中，由于成像速度较慢，限制了其时空分辨率。获取具有较好时空分辨率的心脏序列来捕捉心脏的动态活动是一项艰巨的任务。压缩感知（CS）理论已被用于提高成像速度，从而提高时空分辨率。本研究计划的另一个主要目的是通过利用时空稀疏性和有效的螺旋轨迹来改善欠采样数据的CS重建。此外，在心脏灌注和心脏影像对应的高加速动态MRI数据重建中，主要挑战是该方法对大帧间运动的敏感性。具体来说，信号表示的紧凑性随着帧间运动而降低，从而限制了最大可能的加速度。在较高的加速度因素下，重建过程往往会出现时间模糊和运动相关的伪影。为了解决上述挑战，将开发几种运动估计和补偿方案。我们的目标是解决运动伪影的问题，包括呼吸和心脏运动在压缩传感重建。计算机断层扫描（CT）被认为是一种高剂量的应用，对患者的健康有害，特别是在重复随访和儿科中。为了降低曝光效应，可以降低成像剂量，从而降低信噪比。本研究计划的第二部分将侧重于开发信号处理技术，用于利用稀疏表示对低剂量CT图像进行去噪。本文将提出不同的稀疏编码和字典学习技术。将研究深度学习和强化学习方法，并开发新的算法来直接从正常剂量的CT图像中学习。通过从低剂量和正常剂量图像中学习各种低水平到高水平的特征，我们将从低剂量图像中生成高质量的无噪声CT图像。我们正在与萨斯喀彻温大学医学影像系合作。

Application of DSP in medical image processing has become an integral part of nearly every procedure, machine and system utilized in medical centers today. The overall objective of the proposed research program is to develop novel signal and image processing algorithms and implement robust and reliable techniques for medical imaging systems. Deep convolutional neural networks and compressed sensing approaches will be investigated throughout this research. The fundamental and theoretical elements of the research include image analysis, formation, enhancement, registration, reconstruction and modeling. In dynamic cardiac cine Magnetic Resonance Imaging (MRI), scan times are lengthy, because the spatial frequency information is retrieved sequentially and the image will be formed only after all the frequency information is gathered. Therefore, in these acquired images, the spatio-temporal resolution is limited by the low imaging speed. It is a demanding task to obtain cardiac sequences with better spatio-temporal resolution to capture the dynamic activity of the heart. Compressed sensing (CS) theory has been applied to improve the imaging speed and thus the spatio-temporal resolution. Another main purpose of this research proposal is to improve CS reconstruction of under-sampled data by exploiting spatio-temporal sparsity and efficient spiral trajectories. In addition, in the reconstruction of highly accelerated dynamic MRI data corresponding to cardiac perfusion and cardiac cine, the main challenge is the sensitivity of the methods to large inter-frame motion. Specifically, the compactness of the signal representation decreases with inter-frame motion, thus restricting the maximum possible acceleration. The reconstructions often suffer from temporal blurring and motion-related artifacts at higher acceleration factors. To address the above challenge, several motion-estimation and compensation schemes will be developed. We aim to address the problem of motion artifacts including respiratory and cardiac motion in compressed sensing reconstructions. Computed Tomography (CT) is considered to be a high-dose application and is harmful to patient's health specially in repeated follow ups and in pediatrics. To reduce the exposure effect, the imaging dosage can be lowered, which leads to reduced signal to noise ratio. The second part of this research proposal will focus on developing signal processing techniques for denoising low-dose CT images, utilizing the sparse representations. Different sparse coding and dictionary learning techniques will be proposed. Deep learning and reinforcement learning approaches will be investigated and novel algorithms will be developed to learn directly from the normal-dose CT images. By learning various low-level to high-level features from low-dose and normal-dose images, we will generate high-quality noise free CT images from low-dose images. We are collaborating with the Dept. of Medical Imaging, Univ. of Saskatchewan.