博士论文答辩：x射线CT图像重建的混合监督-无监督学习框架
日期：2024/05/22 - 2024/05/22
博士论文答辩：x射线CT图像重建的混合监督-无监督学习框架
主讲人：Ling Chen
时间：2024年5月22日（周三）上午9:30
地点：龙宾楼415A会议室
讲座摘要
Computed tomography (CT) imaging is a commonly utilized diagnostic tool in the medical field that provides detailed images of the body’s internal structures. However, traditional CT exposes patients to high levels of radiation, which may pose risks. Low-dose CT (LDCT) has been developed to reduce patient radiation exposure. In this dissertation, several methods have been proposed for improving low-dose CT image reconstruction by leveraging both supervised and unsupervised learned models.
The first method combines penalized weighted least squares (PWLS) reconstruction with regularization based on a union of undercomplete (and alternatively overcomplete) sparsifying transforms learned from datasets. This method is capable of learning sparsifying transforms that are either undercomplete or overcomplete, in contrast to earlier methods such as sparsifying transform (ST) and a union of learned transforms (ULTRA), which were restricted to learning square sparsifying transforms. The union of undercomplete or overcomplete sparsifying transforms are learned in an unsupervised fashion using the conjugate gradient algorithm. The proposed cost function for reconstruction can be effectively addressed by alternately executing an image update step, followed by a step of sparse coding and clustering. Testing on the XCAT phantom demonstrates that this method noticeably enhances the quality of the reconstructed images when compared to the PWLS reconstruction that uses a nonadaptive edge-preserving (EP) regularizer, and the reconstruction results obtained with the learned union of overcomplete sparsifying transforms show improved image details compared to those obtained using ULTRA.
The second method is a network-structured sparsifying transform learning approach for X-ray CT, called multi-layer clustering-based residual sparsifying transform (MCST) learning. The proposed MCST scheme learns multiple different unitary transforms in each layer by dividing each layer’s input into several classes. Unlike the previous multi-layer residual sparsifying transform (MARS) model, which is limited to learning a single sparsifying transform at each layer, and the earlier ULTRA model, which can only learn a union of sparsifying transforms in the initial image domain, the MCST model has an enhanced capability. It can learn multiple unions of sparsifying transforms across both the image domain of the first layer and the residual domains of subsequent layers. The MCST model is trained in an unsupervised manner via a block coordinate descent algorithm. The MCST model is then deployed into the regularizer in PWLS reconstruction for LDCT reconstruction, resulting in better image reconstruction quality than the conventional filtered backprojection (FBP) method and PWLS-EP method. It also outperforms recent advanced methods like PWLS with MARS
and PWLS with ULTRA, especially for displaying clear edges and preserving subtle details.
The third method, WavResNet, is a supervised learning approach for image reconstruction. It operates by estimating the noise levels within the wavelet transform of each input image. By subtracting the estimated noise from the wavelet coefficients, WavResNet effectively denoises the input images. Initially applied to LDCT reconstruction, WavResNet shows superior performance over traditional FBP and model-based image reconstruction (MBIR) methods. To evaluate its generalization, a new method based on WavResNet is also proposed for lensless inline holographic microscopy (LIHM) image enhancement. This method denoises the image in the wavelet domain, different from other supervised learning methods that denoise the image in the image domain.It successfully reduces twin image artifacts and enhances the resolution of the reconstructed images compared to traditional techniques.
The fourth method is a hybrid supervised-unsupervised (SUPER) learning framework for X-ray CT image reconstruction. A general learning framework is proposed that combines a learning-based unsupervised solver and supervised trained neural network reconstructors to simulate a fixed-point iteration process. The general learning framework is first proposed to combine different supervised and unsupervised learning methods. From it, we can derive not only the recent unrolling methods and serial SUPER, which combines the supervised and unsupervised modules sequentially, but also some novel frameworks. Particularly, we extensively explore a parallel SUPER learning framework. This model merges the unsupervised and supervised modules in a parallel configuration, with various unsupervised modules being tested within the parallel SUPER approach. The results demonstrate higher image quality of the proposed framework compared to the standalone supervised method, the standalone unsupervised method and the serial SUPER method.
Overall, this dissertation introduces four distinct machine learning methods designed to improve LDCT image reconstruction. Two novel methods are based on unsupervised learning, one method employs a supervised learning approach, and the fourth method establishes a framework integrating both supervised and unsupervised methods. These advances in machine learning for LDCT image reconstruction improve image quality and have the potential to significantly enhance the accuracy of clinical medical diagnosis.
主讲人简介：
Ling Chen received the B.S. degree in Communication engineering from Huazhong University of Science & Technology in 2018. He is currently a Ph.D. candidate at the UM-SJTU Joint Institute, supervised by Prof. Yong Long. His current research interests include CT image reconstruction, low-dose CT, and machine learning.