Split-spectrum amplitude-decorrelation angiography with optical coherence tomography

Amplitude decorrelation measurement is sensitive to transverse flow and immune to phase noise in comparison to Doppler and other phase-based approaches. However, the high axial resolution of OCT makes it very sensitive to the pulsatile bulk motion noise in the axial direction. To overcome this limitation, we developed split-spectrum amplitude-decorrelation angiography (SSADA) to improve the signal-to-noise ratio (SNR) of flow detection. The full OCT spectrum was split into several narrower bands. Inter-B-scan decorrelation was computed using the spectral bands separately and then averaged. The SSADA algorithm was tested on in vivo images of the human macula and optic nerve head. It significantly improved both SNR for flow detection and connectivity of microvascular network when compared to other amplitude-decorrelation algorithms.

Pooling across cells to normalize single-cell RNA sequencing data with many zero counts

Normalization of single-cell RNA sequencing data is necessary to eliminate cell-specific biases prior to downstream analyses. However, this is not straightforward for noisy single-cell data where many counts are zero. We present a novel approach where expression values are summed across pools of cells, and the summed values are used for normalization. Pool-based size factors are then deconvolved to yield cell-based factors. Our deconvolution approach outperforms existing methods for accurate normalization of cell-specific biases in simulated data. Similar behavior is observed in real data, where deconvolution improves the relevance of results of downstream analyses.

3D variability analysis: Resolving continuous flexibility and discrete heterogeneity from single particle cryo-EM

Single particle cryo-EM excels in determining static structures of protein molecules, but existing 3D reconstruction methods have been ineffective in modelling flexible proteins. We introduce 3D variability analysis (3DVA), an algorithm that fits a linear subspace model of conformational change to cryo-EM data at high resolution. 3DVA enables the resolution and visualization of detailed molecular motions of both large and small proteins, revealing new biological insight from single particle cryo-EM data. Experimental results demonstrate the ability of 3DVA to resolve multiple flexible motions of α-helices in the sub-50 kDa transmembrane domain of a GPCR complex, bending modes of a sodium ion channel, five types of symmetric and symmetry-breaking flexibility in a proteasome, large motions in a spliceosome complex, and discrete conformational states of a ribosome assembly. 3DVA is implemented in the cryoSPARC software package.

Comprehensive evaluation of differential gene expression analysis methods for RNA-seq data

A large number of computational methods have been developed for analyzing differential gene expression in RNA-seq data. We describe a comprehensive evaluation of common methods using the SEQC benchmark dataset and ENCODE data. We consider a number of key features, including normalization, accuracy of differential expression detection and differential expression analysis when one condition has no detectable expression. We find significant differences among the methods, but note that array-based methods adapted to RNA-seq data perform comparably to methods designed for RNA-seq. Our results demonstrate that increasing the number of replicate samples significantly improves detection power over increased sequencing depth.

Generative Adversarial Networks for Noise Reduction in Low-Dose CT

Noise is inherent to low-dose CT acquisition. We propose to train a convolutional neural network (CNN) jointly with an adversarial CNN to estimate routine-dose CT images from low-dose CT images and hence reduce noise. A generator CNN was trained to transform low-dose CT images into routine-dose CT images using voxelwise loss minimization. An adversarial discriminator CNN was simultaneously trained to distinguish the output of the generator from routine-dose CT images. The performance of this discriminator was used as an adversarial loss for the generator. Experiments were performed using CT images of an anthropomorphic phantom containing calcium inserts, as well as patient non-contrast-enhanced cardiac CT images. The phantom and patients were scanned at 20% and 100% routine clinical dose. Three training strategies were compared: the first used only voxelwise loss, the second combined voxelwise loss and adversarial loss, and the third used only adversarial loss. The results showed that training with only voxelwise loss resulted in the highest peak signal-to-noise ratio with respect to reference routine-dose images. However, CNNs trained with adversarial loss captured image statistics of routine-dose images better. Noise reduction improved quantification of low-density calcified inserts in phantom CT images and allowed coronary calcium scoring in low-dose patient CT images with high noise levels. Testing took less than 10 s per CT volume. CNN-based low-dose CT noise reduction in the image domain is feasible. Training with an adversarial network improves the CNNs ability to generate images with an appearance similar to that of reference routine-dose CT images.

Deep learning in medical imaging and radiation therapy

The goals of this review paper on deep learning (DL) in medical imaging and radiation therapy are to (a) summarize what has been achieved to date; (b) identify common and unique challenges, and strategies that researchers have taken to address these challenges; and (c) identify some of the promising avenues for the future both in terms of applications as well as technical innovations. We introduce the general principles of DL and convolutional neural networks, survey five major areas of application of DL in medical imaging and radiation therapy, identify common themes, discuss methods for dataset expansion, and conclude by summarizing lessons learned, remaining challenges, and future directions.

Ultrastructural Details of Mammalian Chromosome Architecture

Over the past decade, 3C-related methods have provided remarkable insights into chromosome folding in vivo. To overcome the limited resolution of prior studies, we extend a recently developed Hi-C variant, Micro-C, to map chromosome architecture at nucleosome resolution in human ESCs and fibroblasts. Micro-C robustly captures known features of chromosome folding including compartment organization, topologically associating domains, and interactions between CTCF binding sites. In addition, Micro-C provides a detailed map of nucleosome positions and localizes contact domain boundaries with nucleosomal precision. Compared to Hi-C, Micro-C exhibits an order of magnitude greater dynamic range, allowing the identification of ∼20,000 additional loops in each cell type. Many newly identified peaks are localized along extrusion stripes and form transitive grids, consistent with their anchors being pause sites impeding cohesin-dependent loop extrusion. Our analyses comprise the highest-resolution maps of chromosome folding in human cells to date, providing a valuable resource for studies of chromosome organization.

Photon-counting Detector CT: System Design and Clinical Applications of an Emerging Technology

Photon-counting detector (PCD) CT is an emerging technology that has shown tremendous progress in the last decade. Various types of PCD CT systems have been developed to investigate the benefits of this technology, which include reduced electronic noise, increased contrast-to-noise ratio with iodinated contrast material and radiation dose efficiency, reduced beam-hardening and metal artifacts, extremely high spatial resolution (33 line pairs per centimeter), simultaneous multienergy data acquisition, and the ability to image with and differentiate among multiple CT contrast agents. PCD technology is described and compared with conventional CT detector technology. With the use of a whole-body research PCD CT system as an example, PCD technology and its use for in vivo high-spatial-resolution multienergy CT imaging is discussed. The potential clinical applications, diagnostic benefits, and challenges associated with this technology are then discussed, and examples with phantom, animal, and patient studies are provided. ©RSNA, 2019.

Highly skin-conformal microhairy sensor for pulse signal amplification

A bioinspired microhairy sensor is developed to enable ultraconformability on nonflat surfaces and significant enhancement in the signal-to-noise ratio of the retrieved signals. The device shows ≈12 times increase in the signal-to-noise ratio in the generated capacitive signals, allowing the ultraconformal microhair pressure sensors to be capable of measuring weak pulsations of internal jugular venous pulses stemming from a human neck.

Determination of the appropriate b value and number of gradient directions for high-angular-resolution diffusion-weighted imaging

High-angular-resolution diffusion-weighted imaging (HARDI) is one of the most common MRI acquisition schemes for use with higher order models of diffusion. However, the optimal b value and number of diffusion-weighted (DW) directions for HARDI are still undetermined, primarily as a result of the large number of available reconstruction methods and corresponding parameters, making it impossible to identify a single criterion by which to assess performance. In this study, we estimate the minimum number of DW directions and optimal b values required for HARDI by focusing on the angular frequency content of the DW signal itself. The spherical harmonic (SH) series provides the spherical analogue of the Fourier series, and can hence be used to examine the angular frequency content of the DW signal. Using high-quality data acquired along 500 directions over a range of b values, we estimate that SH terms above l = 8 are negligible in practice for b values up to 5000 s/mm(2), implying that a minimum of 45 DW directions is sufficient to fully characterise the DW signal. l > 0 SH terms were found to increase as a function of b value, levelling off at b = 3000 s/mm(2), suggesting that this value already provides the highest achievable angular resolution. In practice, it is recommended to acquire more than the minimum of 45 DW directions to avoid issues with imperfections in the uniformity of the DW gradient directions and to meet signal-to-noise requirements of the intended reconstruction method.

Computational segmentation of collagen fibers from second-harmonic generation images of breast cancer

Second-harmonic generation (SHG) imaging can help reveal interactions between collagen fibers and cancer cells. Quantitative analysis of SHG images of collagen fibers is challenged by the heterogeneity of collagen structures and low signal-to-noise ratio often found while imaging collagen in tissue. The role of collagen in breast cancer progression can be assessed post acquisition via enhanced computation. To facilitate this, we have implemented and evaluated four algorithms for extracting fiber information, such as number, length, and curvature, from a variety of SHG images of collagen in breast tissue. The image-processing algorithms included a Gaussian filter, SPIRAL-TV filter, Tubeness filter, and curvelet-denoising filter. Fibers are then extracted using an automated tracking algorithm called fiber extraction (FIRE). We evaluated the algorithm performance by comparing length, angle and position of the automatically extracted fibers with those of manually extracted fibers in twenty-five SHG images of breast cancer. We found that the curvelet-denoising filter followed by FIRE, a process we call CT-FIRE, outperforms the other algorithms under investigation. CT-FIRE was then successfully applied to track collagen fiber shape changes over time in an in vivo mouse model for breast cancer.

Advances in analysis of low signal-to-noise images link dynamin and AP2 to the functions of an endocytic checkpoint

Numerous endocytic accessory proteins (EAPs) mediate assembly and maturation of clathrin-coated pits (CCPs) into cargo-containing vesicles. Analysis of EAP function through bulk measurement of cargo uptake has been hampered due to potential redundancy among EAPs and, as we show here, the plasticity and resilience of clathrin-mediated endocytosis (CME). Instead, EAP function is best studied by uncovering the correlation between variations in EAP association to individual CCPs and the resulting variations in maturation. However, most EAPs bind to CCPs in low numbers, making the measurement of EAP association via fused fluorescent reporters highly susceptible to detection errors. Here, we present a framework for unbiased measurement of EAP recruitment to CCPs and their direct effects on CCP dynamics. We identify dynamin and the EAP-binding α-adaptin appendage domain of the AP2 adaptor as switches in a regulated, multistep maturation process and provide direct evidence for a molecular checkpoint in CME.

Diffusion weighted image denoising using overcomplete local PCA

Diffusion Weighted Images (DWI) normally shows a low Signal to Noise Ratio (SNR) due to the presence of noise from the measurement process that complicates and biases the estimation of quantitative diffusion parameters. In this paper, a new denoising methodology is proposed that takes into consideration the multicomponent nature of multi-directional DWI datasets such as those employed in diffusion imaging. This new filter reduces random noise in multicomponent DWI by locally shrinking less significant Principal Components using an overcomplete approach. The proposed method is compared with state-of-the-art methods using synthetic and real clinical MR images, showing improved performance in terms of denoising quality and estimation of diffusion parameters.

Topaz-Denoise: general deep denoising models for cryoEM and cryoET

Cryo-electron microscopy (cryoEM) is becoming the preferred method for resolving protein structures. Low signal-to-noise ratio (SNR) in cryoEM images reduces the confidence and throughput of structure determination during several steps of data processing, resulting in impediments such as missing particle orientations. Denoising cryoEM images can not only improve downstream analysis but also accelerate the time-consuming data collection process by allowing lower electron dose micrographs to be used for analysis. Here, we present Topaz-Denoise, a deep learning method for reliably and rapidly increasing the SNR of cryoEM images and cryoET tomograms. By training on a dataset composed of thousands of micrographs collected across a wide range of imaging conditions, we are able to learn models capturing the complexity of the cryoEM image formation process. The general model we present is able to denoise new datasets without additional training. Denoising with this model improves micrograph interpretability and allows us to solve 3D single particle structures of clustered protocadherin, an elongated particle with previously elusive views. We then show that low dose collection, enabled by Topaz-Denoise, improves downstream analysis in addition to reducing data collection time. We also present a general 3D denoising model for cryoET. Topaz-Denoise and pre-trained general models are now included in Topaz. We expect that Topaz-Denoise will be of broad utility to the cryoEM community for improving micrograph and tomogram interpretability and accelerating analysis.

On the definition of signal-to-noise ratio and contrast-to-noise ratio for FMRI data

Signal-to-noise ratio, the ratio between signal and noise, is a quantity that has been well established for MRI data but is still subject of ongoing debate and confusion when it comes to fMRI data. fMRI data are characterised by small activation fluctuations in a background of noise. Depending on how the signal of interest and the noise are identified, signal-to-noise ratio for fMRI data is reported by using many different definitions. Since each definition comes with a different scale, interpreting and comparing signal-to-noise ratio values for fMRI data can be a very challenging job. In this paper, we provide an overview of existing definitions. Further, the relationship with activation detection power is investigated. Reference tables and conversion formulae are provided to facilitate comparability between fMRI studies.

Opportunities in Interventional and Diagnostic Imaging by Using High-Performance Low-Field-Strength MRI

Background Commercial low-field-strength MRI systems are generally not equipped with state-of-the-art MRI hardware, and are not suitable for demanding imaging techniques. An MRI system was developed that combines low field strength (0.55 T) with high-performance imaging technology. Purpose To evaluate applications of a high-performance low-field-strength MRI system, specifically MRI-guided cardiovascular catheterizations with metallic devices, diagnostic imaging in high-susceptibility regions, and efficient image acquisition strategies. Materials and Methods A commercial 1.5-T MRI system was modified to operate at 0.55 T while maintaining high-performance hardware, shielded gradients (45 mT/m; 200 T/m/sec), and advanced imaging methods. MRI was performed between January 2018 and April 2019. T1, T2, and T2* were measured at 0.55 T; relaxivity of exogenous contrast agents was measured; and clinical applications advantageous at low field were evaluated. Results There were 83 0.55-T MRI examinations performed in study participants (45 women; mean age, 34 years ± 13). On average, T1 was 32% shorter, T2 was 26% longer, and T2* was 40% longer at 0.55 T compared with 1.5 T. Nine metallic interventional devices were found to be intrinsically safe at 0.55 T (<1°C heating) and MRI-guided right heart catheterization was performed in seven study participants with commercial metallic guidewires. Compared with 1.5 T, reduced image distortion was shown in lungs, upper airway, cranial sinuses, and intestines because of improved field homogeneity. Oxygen inhalation generated lung signal enhancement of 19% ± 11 (standard deviation) at 0.55 T compared with 7.6% ± 6.3 at 1.5 T (P = .02; five participants) because of the increased T1 relaxivity of oxygen (4.7e-4 mmHg-1sec-1). Efficient spiral image acquisitions were amenable to low field strength and generated increased signal-to-noise ratio compared with Cartesian acquisitions (P < .02). Representative imaging of the brain, spine, abdomen, and heart generated good image quality with this system. Conclusion This initial study suggests that high-performance low-field-strength MRI offers advantages for MRI-guided catheterizations with metal devices, MRI in high-susceptibility regions, and efficient imaging. © RSNA, 2019 Online supplemental material is available for this article. See also the editorial by Grist in this issue.

A genetically encoded fluorescent acetylcholine indicator for in vitro and in vivo studies

The neurotransmitter acetylcholine (ACh) regulates a diverse array of physiological processes throughout the body. Despite its importance, cholinergic transmission in the majority of tissues and organs remains poorly understood owing primarily to the limitations of available ACh-monitoring techniques. We developed a family of ACh sensors (GACh) based on G-protein-coupled receptors that has the sensitivity, specificity, signal-to-noise ratio, kinetics and photostability suitable for monitoring ACh signals in vitro and in vivo. GACh sensors were validated with transfection, viral and/or transgenic expression in a dozen types of neuronal and non-neuronal cells prepared from multiple animal species. In all preparations, GACh sensors selectively responded to exogenous and/or endogenous ACh with robust fluorescence signals that were captured by epifluorescence, confocal, and/or two-photon microscopy. Moreover, analysis of endogenous ACh release revealed firing-pattern-dependent release and restricted volume transmission, resolving two long-standing questions about central cholinergic transmission. Thus, GACh sensors provide a user-friendly, broadly applicable tool for monitoring cholinergic transmission underlying diverse biological processes.

Ion mobility-derived collision cross section as an additional measure for lipid fingerprinting and identification

Despite recent advances in analytical and computational chemistry, lipid identification remains a significant challenge in lipidomics. Ion-mobility spectrometry provides an accurate measure of the molecules' rotationally averaged collision cross-section (CCS) in the gas phase and is thus related to ionic shape. Here, we investigate the use of CCS as a highly specific molecular descriptor for identifying lipids in biological samples. Using traveling wave ion mobility mass spectrometry (MS), we measured the CCS values of over 200 lipids within multiple chemical classes. CCS values derived from ion mobility were not affected by instrument settings or chromatographic conditions, and they were highly reproducible on instruments located in independent laboratories (interlaboratory RSD < 3% for 98% of molecules). CCS values were used as additional molecular descriptors to identify brain lipids using a variety of traditional lipidomic approaches. The addition of CCS improved the reproducibility of analysis in a liquid chromatography-MS workflow and maximized the separation of isobaric species and the signal-to-noise ratio in direct-MS analyses (e.g., "shotgun" lipidomics and MS imaging). These results indicate that adding CCS to databases and lipidomics workflows increases the specificity and selectivity of analysis, thus improving the confidence in lipid identification compared to traditional analytical approaches. The CCS/accurate-mass database described here is made publicly available.

Short separation channel location impacts the performance of short channel regression in NIRS

Near-Infrared Spectroscopy (NIRS) allows the recovery of cortical oxy- and deoxyhemoglobin changes associated with evoked brain activity. NIRS is a back-reflection measurement making it very sensitive to the superficial layers of the head, i.e. the skin and the skull, where systemic interference occurs. As a result, the NIRS signal is strongly contaminated with systemic interference of superficial origin. A recent approach to overcome this problem has been the use of additional short source-detector separation optodes as regressors. Since these additional measurements are mainly sensitive to superficial layers in adult humans, they can be used to remove the systemic interference present in longer separation measurements, improving the recovery of the cortical hemodynamic response function (HRF). One question that remains to answer is whether or not a short separation measurement is required in close proximity to each long separation NIRS channel. Here, we show that the systemic interference occurring in the superficial layers of the human head is inhomogeneous across the surface of the scalp. As a result, the improvement obtained by using a short separation optode decreases as the relative distance between the short and the long measurement is increased. NIRS data was acquired on 6 human subjects both at rest and during a motor task consisting of finger tapping. The effect of distance between the short and the long channel was first quantified by recovering a synthetic hemodynamic response added over the resting-state data. The effect was also observed in the functional data collected during the finger tapping task. Together, these results suggest that the short separation measurement must be located as close as 1.5 cm from the standard NIRS channel in order to provide an improvement which is of practical use. In this case, the improvement in Contrast-to-Noise Ratio (CNR) compared to a standard General Linear Model (GLM) procedure without using any small separation optode reached 50% for HbO and 100% for HbR. Using small separations located farther than 2 cm away resulted in mild or negligible improvements only.

Super-resolution musculoskeletal MRI using deep learning

PURPOSE

To develop a super-resolution technique using convolutional neural networks for generating thin-slice knee MR images from thicker input slices, and compare this method with alternative through-plane interpolation methods.

METHODS

We implemented a 3D convolutional neural network entitled DeepResolve to learn residual-based transformations between high-resolution thin-slice images and lower-resolution thick-slice images at the same center locations. DeepResolve was trained using 124 double echo in steady-state (DESS) data sets with 0.7-mm slice thickness and tested on 17 patients. Ground-truth images were compared with DeepResolve, clinically used tricubic interpolation, and Fourier interpolation methods, along with state-of-the-art single-image sparse-coding super-resolution. Comparisons were performed using structural similarity, peak SNR, and RMS error image quality metrics for a multitude of thin-slice downsampling factors. Two musculoskeletal radiologists ranked the 3 data sets and reviewed the diagnostic quality of the DeepResolve, tricubic interpolation, and ground-truth images for sharpness, contrast, artifacts, SNR, and overall diagnostic quality. Mann-Whitney U tests evaluated differences among the quantitative image metrics, reader scores, and rankings. Cohen's Kappa (κ) evaluated interreader reliability.

RESULTS

DeepResolve had significantly better structural similarity, peak SNR, and RMS error than tricubic interpolation, Fourier interpolation, and sparse-coding super-resolution for all downsampling factors (p < .05, except 4 × and 8 × sparse-coding super-resolution downsampling factors). In the reader study, DeepResolve significantly outperformed (p < .01) tricubic interpolation in all image quality categories and overall image ranking. Both readers had substantial scoring agreement (κ = 0.73).

CONCLUSION

DeepResolve was capable of resolving high-resolution thin-slice knee MRI from lower-resolution thicker slices, achieving superior quantitative and qualitative diagnostic performance to both conventionally used and state-of-the-art methods.

Progress and Trends in AIE-Based Bioprobes: A Brief Overview

Luminescent bioprobes are powerful analytical means for biosensing and optical imaging. Luminogens featured with aggregation-induced emission (AIE) attributes have emerged as ideal building blocks for high-performance bioprobes. Bioprobes constructed with AIE luminogens have been identified to be a novel class of FL light-up probing tools. In contrast to conventional bioprobes based on the luminophores with aggregation-caused quenching (ACQ) effect, the AIE-based bioprobes enjoy diverse superiorities, such as lower background, higher signal-to-noise ratio and sensitivity, better accuracy, and more outstanding resistance to photobleaching. AIE-based bioprobes have been tailored for a vast variety of purposes ranging from biospecies sensing to bioimaging to theranostics (i.e., image-guided therapies). In this review, recent five years' advances in AIE-based bioprobes are briefly overviewed in a perspective distinct from other reviews, focusing on the most appealing trends and progresses in this flourishing research field. There are altogether 11 trends outlined, which have been classified into four aspects: the probe composition and form (bioconjugtes, nanoprobes), the output signal of probe (far-red/near-infrared luminescence, two/three-photon excited fluorescence, phosphorescence), the modality and functionality of probing system (dual-modality, dual/multifunctionality), the probing object and application outlet (specific organelles, cancer cells, bacteria, real samples). Typical examples of each trend are presented and specifically demonstrated. Some important prospects and challenges are pointed out as well in the hope of intriguing more interests from researchers working in diverse areas into this exciting research field.

Sparse multidimensional iterative lineshape-enhanced (SMILE) reconstruction of both non-uniformly sampled and conventional NMR data

Implementation of a new algorithm, SMILE, is described for reconstruction of non-uniformly sampled two-, three- and four-dimensional NMR data, which takes advantage of the known phases of the NMR spectrum and the exponential decay of underlying time domain signals. The method is very robust with respect to the chosen sampling protocol and, in its default mode, also extends the truncated time domain signals by a modest amount of non-sampled zeros. SMILE can likewise be used to extend conventional uniformly sampled data, as an effective multidimensional alternative to linear prediction. The program is provided as a plug-in to the widely used NMRPipe software suite, and can be used with default parameters for mainstream application, or with user control over the iterative process to possibly further improve reconstruction quality and to lower the demand on computational resources. For large data sets, the method is robust and demonstrated for sparsities down to ca 1%, and final all-real spectral sizes as large as 300 Gb. Comparison between fully sampled, conventionally processed spectra and randomly selected NUS subsets of this data shows that the reconstruction quality approaches the theoretical limit in terms of peak position fidelity and intensity. SMILE essentially removes the noise-like appearance associated with the point-spread function of signals that are a default of five-fold above the noise level, but impacts the actual thermal noise in the NMR spectra only minimally. Therefore, the appearance and interpretation of SMILE-reconstructed spectra is very similar to that of fully sampled spectra generated by Fourier transformation.

A Benchmark Dataset for SSVEP-Based Brain-Computer Interfaces

This paper presents a benchmark steady-state visual evoked potential (SSVEP) dataset acquired with a 40-target brain- computer interface (BCI) speller. The dataset consists of 64-channel Electroencephalogram (EEG) data from 35 healthy subjects (8 experienced and 27 naïve) while they performed a cue-guided target selecting task. The virtual keyboard of the speller was composed of 40 visual flickers, which were coded using a joint frequency and phase modulation (JFPM) approach. The stimulation frequencies ranged from 8 Hz to 15.8 Hz with an interval of 0.2 Hz. The phase difference between two adjacent frequencies was . For each subject, the data included six blocks of 40 trials corresponding to all 40 flickers indicated by a visual cue in a random order. The stimulation duration in each trial was five seconds. The dataset can be used as a benchmark dataset to compare the methods for stimulus coding and target identification in SSVEP-based BCIs. Through offline simulation, the dataset can be used to design new system diagrams and evaluate their BCI performance without collecting any new data. The dataset also provides high-quality data for computational modeling of SSVEPs. The dataset is freely available fromhttp://bci.med.tsinghua.edu.cn/download.html.