This paper proposes a deep learning approach for accelerating magnetic resonance imaging (MRI) using a large number of existing high quality MR images as the training datasets. An off-line convolutional neural network is designed and trained to identify the mapping relationship between the MR images obtained from zero-filled and fully-sampled k-space data. The network is not only capable of restoring fine structures and details but is also compatible with online constrained reconstruction methods. Experimental results on real MR data have shown encouraging performance of the proposed method for efficient and accurate imaging.

/pdf/accelerating-magnetic-resonance-imaging-via-deep-learning-4d6wbyobwu.pdf

Accelerating magnetic resonance imaging via deep learning

We propose an Iterative Mean Filter (IMF) to eliminate the salt-and-pepper noise. IMF uses the mean of gray values of noise-free pixels in a fixed-size window. Unlike other nonlinear filters, IMF does not enlarge the window size. A large size reduces the accuracy of noise removal. Therefore, IMF only uses a window with a size of $3\times3$ . This feature is helpful for IMF to be able to more precisely evaluate a new gray value for the center pixel. To process high-density noise effectively, we propose an iterative procedure for IMF. In the experiments, we operationalize Peak Signal-to-Noise Ratio (PSNR), Visual Information Fidelity, Image Enhancement Factor, Structural Similarity (SSIM), and Multiscale Structure Similarity to assess image quality. Furthermore, we compare denoising results of IMF with ones of the other state-of-the-art methods. A comprehensive comparison of execution time is also provided. The qualitative results by PSNR and SSIM showed that IMF outperforms the other methods such as Based-on Pixel Density Filter (BPDF), Decision-Based Algorithm (DBA), Modified Decision-Based Untrimmed Median Filter (MDBUTMF), Noise Adaptive Fuzzy Switching Median Filter (NAFSMF), Adaptive Weighted Mean Filter (AWMF), Different Applied Median Filter (DAMF), Adaptive Type-2 Fuzzy Filter (FDS): for the IMAGESTEST dataset - BPDF (25.36/0.756), DBA (28.72/0.8426), MDBUTMF (25.93/0.8426), NAFSMF (29.32/0.8735), AWMF (32.25/0.9177), DAMF (31.65/0.9154), FDS (27.98/0.8338), and IMF (33.67/0.9252); and for the BSDS dataset - BPDF (24.95/0.7469), DBA (26.84/0.8061), MDBUTMF (26.25/0.7732), NAFSMF (27.26/0.8191), AWMF (28.89/0.8672), DAMF (29.11/0.8667), FDS (26.85/0.8095), and IMF (30.04/0.8753).

/pdf/an-iterative-mean-filter-for-image-denoising-31athaae9w.pdf

An Iterative Mean Filter for Image Denoising

The term “blind denoising” refers to the fact that the basis used for denoising is learned from the noisy sample itself during denoising. Dictionary learning- and transform learning-based formulations for blind denoising are well known. But there has been no autoencoder-based solution for the said blind denoising approach. So far, autoencoder-based denoising formulations have learned the model on a separate training data and have used the learned model to denoise test samples. Such a methodology fails when the test image (to denoise) is not of the same kind as the models learned with. This will be the first work, where we learn the autoencoder from the noisy sample while denoising. Experimental results show that our proposed method performs better than dictionary learning (K-singular value decomposition), transform learning, sparse stacked denoising autoencoder, and the gold standard BM3D algorithm.

Blind Denoising Autoencoder

A new adaptive switching median filter for impulse noise reduction with pre-detection based on evidential reasoning

In this study, we propose a new method, i.e. Adaptive Riesz Mean Filter (ARmF), by operationalizing pixel similarity for salt-and-pepper noise (SPN) removal. Afterwards, we compare the results of ARmF, A New Adaptive Weighted Mean Filter (AWMF), Different Applied Median Filter (DAMF), Noise Adaptive Fuzzy Switching Median Filter (NAFSMF), Based on Pixel Density Filter (BPDF), Modified Decision-Based Unsymmetric Trimmed Median Filter (MDBUTMF) and Decision-Based Algorithm (DBA) by using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Image Enhancement Factor (IEF), and Visual Information Fidelity (VIF) for 20 traditional test images (Lena, Cameraman, Barbara, Baboon, Peppers, Living Room, Lake, Plane, Hill, Pirate, Boat, House, Bridge, Elaine, Flintstones, Flower, Parrot, Dark-Haired Woman, Blonde Woman, and Einstein), 40 test images in the TESTIMAGES Database, and 200 RGB test images from the UC Berkeley Dataset ranging in noise density from 10% to 90%. Moreover, we compare the running time of these algorithms. These results show that ARmF outperforms the methods mentioned above. We finally discuss the need for further research.

Pixel similarity-based adaptive Riesz mean filter for salt-and-pepper noise removal

Dictionary learning based impulse noise removal via L1-L1 minimization

This paper proposes a spatially adaptive constrained dictionary learning (SAC-DL) algorithm for Rician noise removal in magnitude magnetic resonance (MR) images. This algorithm explores both the strength of dictionary learning to preserve image structures and the robustness of local variance estimation to remove signal-dependent Rician noise. The magnitude image is first separated into a number of partly overlapping image patches. The statistics of each patch are collected and analyzed to obtain a local noise variance. To better adapt to Rician noise, a correction factor is formulated with the local signal-to-noise ratio (SNR). Finally, the trained dictionary is used to denoise each image patch under spatially adaptive constraints. The proposed algorithm has been compared to the popular nonlocal means (NLM) filtering and unbiased NLM (UNLM) algorithm on simulated T1-weighted, T2-weighted and PD-weighted MR images. Our results suggest that the SAC-DL algorithm preserves more image structures while effectively removing the noise than NLM and it is also superior to UNLM at low noise levels.

Magnetic resonance image restoration via dictionary learning under spatially adaptive constraints

Due to the imperfections of the radio frequency (RF) coil or object-dependent electrodynamic interactions, magnetic resonance (MR) images often suffer from a smooth and biologically meaningless bias field, which causes severe troubles for subsequent processing and quantitative analysis. To effectively restore the original signal, this paper simultaneously exploits the spatial and gradient features of the corrupted MR images for bias correction via the joint entropy regularization. With both isotropic and anisotropic total variation (TV) considered, two nonparametric bias correction algorithms have been proposed, namely IsoTVBiasC and AniTVBiasC. These two methods have been applied to simulated images under various noise levels and bias field corruption and also tested on real MR data. The test results show that the proposed two methods can effectively remove the bias field and also present comparable performance compared to the state-of-the-art methods.

Bias correction for magnetic resonance images via joint entropy regularization.

Positron Emission Tomography (PET) has become a preferred imaging modality for cancer diagnosis, radiotherapy planning, and treatment responses monitoring. Accurate and automatic tumor segmentation is the fundamental requirement for these clinical applications. Deep convolutional neural networks have become the state-of-the-art in PET tumor segmentation. The normalization process is one of the key components for accelerating network training and improving the performance of the network. However, existing normalization methods either introduce batch noise into the instance PET image by calculating statistics on batch level or introduce background noise into every single pixel by sharing the same learnable parameters spatially. In this paper, we proposed an attentive transformation (AT)-based normalization method for PET tumor segmentation. We exploit the distinguishability of breast tumor in PET images and dynamically generate dedicated and pixel-dependent learnable parameters in normalization via the transformation on a combination of channel-wise and spatial-wise attentive responses. The attentive learnable parameters allow to re-calibrate features pixel-by-pixel to focus on the high-uptake area while attenuating the background noise of PET images. Our experimental results on two real clinical datasets show that the AT-based normalization method improves breast tumor segmentation performance when compared with the existing normalization methods.

Improving Breast Tumor Segmentation in PET via Attentive Transformation Based Normalization

Dynamic PET imaging provides superior physiological information than conventional static PET imaging. However, the dynamic information is gained at the cost of a long scanning protocol; this limits the clinical application of dynamic PET imaging. We developed a modified Logan reference plot model to shorten the acquisition procedure in dynamic PET imaging by omitting the early-time information necessary for the conventional reference Logan model. The proposed model is accurate theoretically, but the straightforward approach raises the sampling problem in implementation and results in noisy parametric images. We then designed a self-supervised convolutional neural network to increase the noise performance of parametric imaging, with dynamic images of only a single subject for training. The proposed method was validated via simulated and real dynamic 18F-fallypride PET data. Results showed that it accurately estimated the distribution volume ratio (DVR) in dynamic PET with a shortened scanning protocol, e.g., 20 minutes, where the estimations were comparable with those obtained from a standard dynamic PET study of 120 minutes of acquisition. Further comparisons illustrated that our method outperformed the shortened Logan model implemented with Gaussian filtering, regularization, BM4D and the 4D deep image prior methods in terms of the trade-off between bias and variance. Since the proposed method uses data acquired in a short period of time upon the equilibrium, it has the potential to add clinical values by providing both DVR and Standard Uptake Value (SUV) simultaneously. It thus promotes clinical applications of dynamic PET studies when neuronal receptor functions are studied.

Qiu Huang

Papers

Dictionary learning based impulse noise removal via L1-L1 minimization

Magnetic resonance image restoration via dictionary learning under spatially adaptive constraints

Bias correction for magnetic resonance images via joint entropy regularization.

Improving Breast Tumor Segmentation in PET via Attentive Transformation Based Normalization

A Shortened Model for Logan Reference Plot Implemented via the Self-Supervised Neural Network for Parametric PET Imaging.