Reconstruction Task Research Articles

Parkinsonian gait modelling from an anomaly deep representation

AbstractParkinson’s Disease (PD) is associated with gait movement disorders, such as bradykinesia, stiffness, tremors and postural instability. Hence, a kinematic gait analysis for PD characterization is key to support diagnosis and to carry out an effective treatment planning. Nowadays, automatic classification and characterization strategies are based on deep learning representations, following supervised rules, and assuming large and stratified data. Nonetheless, such requirements are far from real clinical scenarios. Additionally, supervised rules may introduce bias into architectures from expert’s annotations. This work introduces a self-supervised generative representation to learn gait-motion-related patterns, under the pretext task of video reconstruction. Following an anomaly detection framework, the proposed architecture can avoid inter-class variance, learning hidden and complex kinematics locomotion relationships. In this study, the proposed model was trained and validated with an owner dataset (14 Parkinson and 23 control). Also, an external public dataset (16 Parkinson, 30 control, and 50 Knee-arthritis) was used only for testing, measuring the generalization capability of the method. During training, the method learns from control subjects, while Parkinson subjects are detected as anomaly samples. From owner dataset, the proposed approach achieves a ROC-AUC of 95% in classification task. Regarding the external dataset, the architecture evidence generalization capabilities, achieving a 75% of ROC-AUC (shapeness and homoscedasticity of 66.7%), without any additional training. The proposed model has remarkable performance in detecting gait parkinsonian patterns, recorded in markerless videos, even competitive results with classes non-observed during training.

Deep unfolding network with spatial alignment for multi-modal MRI reconstruction

Multi-modal Magnetic Resonance Imaging (MRI) offers complementary diagnostic information, but some modalities are limited by the long scanning time. To accelerate the whole acquisition process, MRI reconstruction of one modality from highly under-sampled k-space data with another fully-sampled reference modality is an efficient solution. However, the misalignment between modalities, which is common in clinic practice, can negatively affect reconstruction quality. Existing deep learning-based methods that account for inter-modality misalignment perform better, but still share two main common limitations: (1) The spatial alignment task is not adaptively integrated with the reconstruction process, resulting in insufficient complementarity between the two tasks; (2) the entire framework has weak interpretability. In this paper, we construct a novel Deep Unfolding Network with Spatial Alignment, termed DUN-SA, to appropriately embed the spatial alignment task into the reconstruction process. Concretely, we derive a novel joint alignment-reconstruction model with a specially designed aligned cross-modal prior term. By relaxing the model into cross-modal spatial alignment and multi-modal reconstruction tasks, we propose an effective algorithm to solve this model alternatively. Then, we unfold the iterative stages of the proposed algorithm and design corresponding network modules to build DUN-SA with interpretability. Through end-to-end training, we effectively compensate for spatial misalignment using only reconstruction loss, and utilize the progressively aligned reference modality to provide inter-modality prior to improve the reconstruction of the target modality. Comprehensive experiments on four real datasets demonstrate that our method exhibits superior reconstruction performance compared to state-of-the-art methods.

Semi-supervised regression based on Representation Learning for fermentation processes

Biofermentation faces challenges in obtaining real-time quality variables, making it necessary to predict these variables. However, the fermentation process data vary in length and lack sufficient labeled data for model establishment. To solve this problem, this study introduces a framework named RL-SSR(Representation Learning-based Semi-Supervised Regression). First, a data rotation mechanism is designed to address the issue of non-equal-length data. Second, representation learning pre-tasks containing contrastive learning and data reconstruction tasks are implemented to introduce a priori knowledge and numeric features. Finally, the pre-trained model will be fine-tuned with limited labeled data. Experimental results using an industrial-scale penicillin fermentation dataset reveal that RL-SSR outperforms other baseline models, particularly with a small number of labels, confirming the robustness and effectiveness of RL-SSR in the real-time prediction of quality variables in fermentation processes.

Deep-learning-based image reconstruction with limited data: generating synthetic raw data using deep learning.

Deep learning has shown great promise for fast reconstruction of accelerated MRI acquisitions by learning from large amounts of raw data. However, raw data is not always available in sufficient quantities. This study investigates synthetic data generation to complement small datasets and improve reconstruction quality. An adversarial auto-encoder was trained to generate phase and coil sensitivity maps from magnitude images, which were combined into synthetic raw data. On a fourfold accelerated MR reconstruction task, deep-learning-based reconstruction networks were trained with varying amounts of training data (20 to 160 scans). Test set performance was compared between baseline experiments and experiments that incorporated synthetic training data. Training with synthetic raw data showed decreasing reconstruction errors with increasing amounts of training data, but importantly this was magnitude-only data, rather than real raw data. For small training sets, training with synthetic data decreased the mean absolute error (MAE) by up to 7.5%, whereas for larger training sets the MAE increased by up to 2.6%. Synthetic raw data generation improved reconstruction quality in scenarios with limited training data. A major advantage of synthetic data generation is that it allows for the reuse of magnitude-only datasets, which are more readily available than raw datasets.

Drivers’ situational awareness of surrounding vehicles during takeovers: Evidence from a driving simulator study

This study aimed to understand the influence of surrounding vehicle configuration, driving lane, and traffic density on drivers’ situational awareness (SA), takeover performance, and eye-tracking behaviors in conditionally automated driving. An experiment was conducted with the participation of 40 university students using a fixed-base driving simulator configured to simulate SAE Level 3 automation. During the experiment, participants were engaged in playing Tetris on a tablet as a non-driving related task in automated driving mode. Upon hearing an auditory takeover request, participants were instructed to take control of the vehicle, and then complete a scene reconstruction task to report their SA after transferring control back to the automated driving system. Our findings showed that drivers often neglected vehicles at their sides and rear during the takeover, which was associated with higher collision risks. Higher oncoming traffic density led to drivers’ worse SA of surrounding vehicles but more cautious driving behavior. Driving in the right lane generally resulted in smoother takeovers with lower collision risks. Interestingly, while SA did moderate the impacts of driving conditions on safety margins, a higher level of SA did not consistently relate to improved performance, especially in complex scenarios. This suggests the need for support systems that guide drivers to focus on safety–critical objects rather than simply amplifying SA in general. These insights have significant implications for the design of driver monitoring and support systems in automated vehicles.

An efficient dual-domain deep learning network for sparse-view CT reconstruction

Background and ObjectiveWe develop an efficient deep-learning based dual-domain reconstruction method for sparse-view CT reconstruction with small training parameters and comparable running time. We aim to investigate the model's capability and its clinical value by performing objective and subjective quality assessments using clinical CT projection data acquired on commercial scanners. MethodsWe designed two lightweight networks, namely Sino-Net and Img-Net, to restore the projection and image signal from the DD-Net reconstructed images in the projection and image domains, respectively. The proposed network has small training parameters and comparable running time among dual-domain based reconstruction networks and is easy to train (end-to-end). We prospectively collected clinical thoraco-abdominal CT projection data acquired on a Siemens Biograph 128 Edge CT scanner to train and validate the proposed network. Further, we quantitatively evaluated the CT Hounsfield unit (HU) values on 21 organs and anatomic structures, such as the liver, aorta, and ribcage. We also analyzed the noise properties and compared the signal-to-noise ratio (SNR) and the contrast-to-noise ratio (CNR) of the reconstructed images. Besides, two radiologists conducted the subjective qualitative evaluation including the confidence and conspicuity of anatomic structures, and the overall image quality using a 1–5 likert scoring system. ResultsObjective and subjective evaluation showed that the proposed algorithm achieves competitive results in eliminating noise and artifacts, restoring fine structure details, and recovering edges and contours of anatomic structures using 384 views (1/6 sparse rate). The proposed method exhibited good computational cost performance on clinical projection data. ConclusionThis work presents an efficient dual-domain learning network for sparse-view CT reconstruction on raw projection data from a commercial scanner. The study also provides insights for designing an organ-based image quality assessment pipeline for sparse-view reconstruction tasks, potentially benefiting organ-specific dose reduction by sparse-view imaging.

A parametric and feature-based CAD dataset to support human-computer interaction for advanced 3D shape learning

3D shape learning is an important research topic in computer vision, in which the datasets play a critical role. However, most of the existing 3D datasets use voxels, point clouds, mesh, and B-rep, which are not parametric and feature-based. Thus they can not support the generation of real-world engineering computer-aided design (CAD) models with complicated shape features. Furthermore, they are based on 3D geometry results without human-computer interaction (HCI) history. This work is the first to provide a full parametric and feature-based CAD dataset with a selection mechanism to support HCI in 3D learning. First, unlike existing datasets, mainly composed of simple features (typical sketch and extrude), we devise complicated engineering features, such as fillet, chamfer, mirror, pocket, groove, and revolve. Second, different from the monotonous combination of features, we invent a select mechanism to mimic how human focuses on and selects a particular topological entity. The proposed mechanism establishes the relationships among complicated engineering features, which fully express the design intention and design knowledge of human CAD engineers. Therefore, it can process advanced 3D features for real-world engineering shapes. The experiments show that the proposed dataset outperforms existing CAD datasets in both reconstruction and generation tasks. In quantitative experiment, the proposed dataset demonstrates better prediction accuracy than other parametric datasets. Furthermore, CAD models generated from the proposed dataset comply with semantics of the human CAD engineers and can be edited and redesigned via mainstream industrial CAD software.

Reconstruction of infrared absorption and scattering coefficient distributions of semitransparent medium using Adam algorithm combined with adjoint model

Precise reconstruction of infrared optical coefficient distributions within semitransparent medium holds great significance in engineering field. This study introduces the Adam algorithm to reconstruct the infrared absorption and scattering coefficient distributions of a 2D semitransparent slab. The discrete ordinate method (DOM) is utilized to solve the time-domain radiative transfer equation (TD-RTE), and the radiative intensity at boundary is employed as input signal. The gradient calculation of the objective function employs the adjoint equation method (AEM). Results demonstrate that the adaptive moment estimation (Adam) algorithm can simultaneously estimate the infrared absorption and scattering coefficient distributions with high accuracy. Compared to traditional optimization methods, the Adam algorithm demonstrates significant superiority in terms of reconstruction accuracy, computational efficiency, and tolerance to measurement errors. Furthermore, it is also appropriate for determining the infrared optical parameters of different internal structures. Balancing both calculation accuracy and computation time, a recommended learning rate of η = 50 is suggested for the reconstruction task. The Adam algorithm presents limitations for reconstructing the distributions in scattering-dominant medium. The normalized root-mean-square errors of the absorption and scattering coefficients respectively increase to 0.2311 and 0.0655 when the standard measurement error is increased to 20 %.

Интеллектуальная технология оценки остаточного ресурса сложных технических систем

The task of estimating the residual life of complex technical systems has recently become increasingly important. For the pre-trustworthy estimation of this indicator it is required to process large arrays of data on the current state of the system under study. At the same time the task of reconstruction of the model of degradation processes development leading to the occurrence of failures requires solving a number of problems. In this regard, there is a need to use intelligent methods of data processing, which include methods of time series anomaly analysis. Purpose of the study: development of a method for detecting contextual anomalies of time series, allowing to determine the degree of development of degradation processes that lead to the occurrence of failures. Methods. An analogy was established between machine learning methods in supervised, semi-supervised and unsupervised modes and groups of anomaly detection methods differing depending on the degree of availability of labels that characterize the properties and attributes of the corresponding time series, which made it possible to evaluate the features of simple and complex anomalies from unified positions. Results. A spectral method for the analysis of contextual anomalies of time series has been developed, which, in contrast to the known spectral methods involving frequency analysis of time series, uses a special basis of exponential functions; the methodology for calculating the spectral coefficients of the investigated time series, on the basis of which a generalized attribute is calculated, allowing to attribute the investigated case to a normal or anomalous group, is outlined. Conclusion. The proposed method of estimating the residual life of complex technical systems based on the analysis of time series anomalies allows timely detection of the occurrence and development of degradation processes leading to the occurrence of failures, which increases the reliability and safety of technical systems, as well as reduces costs and time for their maintenance.

Hybrid Neural State-Space Modeling for Supervised and Unsupervised Electrocardiographic Imaging.

State-space modeling (SSM) provides a general framework for many image reconstruction tasks. Error in a priori physiological knowledge of the imaging physics, can bring incorrectness to solutions. Modern deep-learning approaches show great promise but lack interpretability and rely on large amounts of labeled data. In this paper, we present a novel hybrid SSM framework for electrocardiographic imaging (ECGI) to leverage the advantage of state-space formulations in data-driven learning. We first leverage the physics-based forward operator to supervise the learning. We then introduce neural modeling of the transition function and the associated Bayesian filtering strategy. We applied the hybrid SSM framework to reconstruct electrical activity on the heart surface from body-surface potentials. In unsupervised settings of both in-silico and in-vivo data without cardiac electrical activity as the ground truth to supervise the learning, we demonstrated improved ECGI performances of the hybrid SSM framework trained from a small number of ECG observations in comparison to the fixed SSM. We further demonstrated that, when in-silico simulation data becomes available, mixed supervised and unsupervised training of the hybrid SSM achieved a further 40.6% and 45.6% improvements, respectively, in comparison to traditional ECGI baselines and supervised data-driven ECGI baselines for localizing the origin of ventricular activations in real data.

An unrolled neural network for accelerated dynamic MRI based on second-order half-quadratic splitting model

The reconstruction of dynamic magnetic resonance images from incomplete k-space data has sparked significant research interest due to its potential to reduce scan time. However, traditional iterative optimization algorithms fail to faithfully reconstruct images at higher acceleration factors and incur long reconstruction time. Furthermore, end-to-end deep learning-based reconstruction algorithms suffer from large model parameters and lack robustness in the reconstruction results. Recently, unrolled deep learning models, have shown immense potential in algorithm stability and applicability flexibility. In this paper, we propose an unrolled deep learning network based on a second-order Half-Quadratic Splitting(HQS) algorithm, where the forward propagation process of this framework strictly follows the computational flow of the HQS algorithm. In particular, we propose a degradation-sense module by associating random sampling patterns with intermediate variables to guide the iterative process. We introduce the Information Fusion Transformer(IFT) to extract both local and non-local prior information from image sequences, thereby removing aliasing artifacts resulting from random undersampling. Finally, we impose low-rank constraints within the HQS algorithm to further enhance the reconstruction results. The experiments demonstrate that each component module of our proposed model contributes to the improvement of the reconstruction task. Our proposed method achieves comparably satisfying performance to the state-of-the-art methods and it exhibits excellent generalization capabilities across different sampling masks. At the low acceleration factor, there is a 0.7% enhancement in the PSNR. Furthermore, when the acceleration factor reached 8 and 12, the PSNR achieves an improvement of 3.4% and 5.8% respectively.

Refining computational inference of gene regulatory networks: integrating knockout data within a multi-task framework.

Constructing accurate gene regulatory network s (GRNs), which reflect the dynamic governing process between genes, is critical to understanding the diverse cellular process and unveiling the complexities in biological systems. With the development of computer sciences, computational-based approaches have been applied to the GRNs inference task. However, current methodologies face challenges in effectively utilizing existing topological information and prior knowledge of gene regulatory relationships, hindering the comprehensive understanding and accurate reconstruction of GRNs. In response, we propose a novel graph neural network (GNN)-based Multi-Task Learning framework for GRN reconstruction, namely MTLGRN. Specifically, we first encode the gene promoter sequences and the gene biological features and concatenate the corresponding feature representations. Then, we construct a multi-task learning framework including GRN reconstruction, Gene knockout predict, and Gene expression matrix reconstruction. With joint training, MTLGRN can optimize the gene latent representations by integrating gene knockout information, promoter characteristics, and other biological attributes. Extensive experimental results demonstrate superior performance compared with state-of-the-art baselines on the GRN reconstruction task, efficiently leveraging biological knowledge and comprehensively understanding the gene regulatory relationships. MTLGRN also pioneered attempts to simulate gene knockouts on bulk data by incorporating gene knockout information.

Low-dimensional signal representations for massive black hole binary signals analysis from LISA data

The space-based gravitational wave observatory LISA will provide a wealth of information to analyze massive black hole binaries with high chirp masses, beyond $10^5$ solar masses. The large number of expected MBHBs (one event a day on average) increases the risk of overlapping between events. As well, the data will be contaminated with non-stationary artifacts, such as glitches and data gaps, which are expected to strongly impact the MBHB analysis, which mandates the development of dedicated detection and retrieval methods on long time intervals. Building upon a methodological approach we introduced for galactic binaries, in this article we investigate an original non-parametric recovery of MBHB signals from measurements with instrumental noise typical of LISA in order to tackle detection and signal reconstruction tasks on long time intervals. We investigated different approaches based on sparse signal modeling and machine learning. In this framework, we focused on recovering MBHB waveforms on long time intervals, which is a building block to further tackling more general signal recovery problems, from gap mitigation to unmixing overlapped signals. To that end, we introduced a hybrid method called SCARF (sparse chirp adaptive representation in Fourier), which combines a deep learning modeling of the merger of the MBHB with a specific adaptive time-frequency representation of the inspiral. Numerical experiments have been carried out on simulations of single MBHB events that account for the LISA response and with realistic realizations of noise. We checked the performances of the proposed hybrid method for the fast detection and recovery of the MBHB.

Unified and Scalable Deep Image Compression Framework for Human and Machine

Image compression aims to minimize the amount of data in image representation while maintaining a certain visual quality for humans, which is an essential technique for storage and transmission. Recently, along with the development of computer vision, machines have become another primary receiver for images and require compressed images at a certain quality level, which may be different from that of human vision. In many scenarios, compressed images should serve both human and machine vision tasks, but few compression methods are designed for both goals simultaneously. In this paper, we propose a unified and scalable deep image compression (USDIC) framework that jointly optimizes the image quality according to human and machine vision in an end-to-end style. For the encoder, we propose an information splitting mechanism (ISM) to separate images into semantic and visual features, which mainly aims at machine analysis and human viewing tasks. For the decoder, we design a scalable decoding architecture. The encoded semantic feature is first decoded for machine analysis tasks, and the image is decoded and reconstructed further by leveraging the decoded semantic features. Herein, to further remove the redundancy between the semantic and visual features of images, we propose a scalable entropy model (SEM) with a joint optimization strategy to reconstruct the image using the two kinds of decoded features. Extensive experimental results show that the proposed USDIC achieves much better performance on the image analysis task while maintaining competitive performance on the traditional image reconstruction task compared with popular image compression methods.

Impact of various DIII-D diagnostics on the accuracy of neural network surrogates for kinetic EFIT reconstructions

Kinetic equilibrium reconstructions make use of profile information such as particle density and temperature measurements in addition to magnetics data to compute a self-consistent equilibrium. They are used in a multitude of physics-based modeling. This work develops a multi-layer perceptron (MLP) neural network (NN) model as a surrogate for kinetic Equilibrium Fitting (EFITs) and trains on the 2019 DIII-D discharge campaign database of kinetic equilibrium reconstructions. We investigate the impact of including various diagnostic data and machine actuator controls as input into the NN. When giving various categories of data as input into NN models that have been trained using those same categories of data, the predictions on multiple equilibrium reconstruction solutions (poloidal magnetic flux, global scalars, pressure profile, current profile) are highly accurate. When comparing different models with different diagnostics as input, the magnetics-only model outputs accurate kinetic profiles and the inclusion of additional data does not significantly impact the accuracy. When the NN is tasked with inferring only a single target such as the EFIT pressure profile or EFIT current profile, we see a large increase in the accuracy of the prediction of the kinetic profiles as more data is included. These results indicate that certain MLP NN configurations can be reasonably robust to different burning-plasma-relevant diagnostics depending on the accuracy requirements for equilibrium reconstruction tasks.

Domain adaptation using AdaBN and AdaIN for high-resolution IVD mesh reconstruction from clinical MRI.

Deep learning has firmly established its dominance in medical imaging applications. However, careful consideration must be exercised when transitioning a trained source model to adapt to an entirely distinct environment that deviates significantly from the training set. The majority of the efforts to mitigate this issue have predominantly focused on classification and segmentation tasks. In this work, we perform a domain adaptation of a trained source model to reconstruct high-resolution intervertebral disc meshes from low-resolution MRI. To address the outlined challenges, we use MRI2Mesh as the shape reconstruction network. It incorporates three major modules: image encoder, mesh deformation, and cross-level feature fusion. This feature fusion module is used to encapsulate local and global disc features. We evaluate two major domain adaptation techniques: adaptive batch normalization (AdaBN) and adaptive instance normalization (AdaIN) for the task of shape reconstruction. Experiments conducted on distinct datasets, including data from different populations, machines, and test sites demonstrate the effectiveness of MRI2Mesh for domain adaptation. MRI2Mesh achieved up to a 14% decrease in Hausdorff distance (HD) and a 19% decrease in the point-to-surface (P2S) metric for both AdaBN and AdaIN experiments, indicating improved performance. MRI2Mesh has demonstrated consistent superiority to the state-of-the-art Voxel2Mesh network across a diverse range of datasets, populations, and scanning protocols, highlighting its versatility. Additionally, AdaBN has emerged as a robust method compared to the AdaIN technique. Further experiments show that MRI2Mesh, when combined with AdaBN, holds immense promise for enhancing the precision of anatomical shape reconstruction in domain adaptation.

Multi-sensor signal fusion for tool wear condition monitoring using denoising transformer auto-encoder Resnet

Multi-sensor signal fusion is commonly used in associate with the artificial intelligence model to monitor tool wears. However, AI models equipped with limited multi-sensor training samples still exist problems: 1) The limitation of training samples may cause the AI model failure due to the intrinsic wear features contaminated by heavy noises. 2) The selection of multi-sensor signals as training samples is a difficult task for the negative impact on the recognition accuracy using inappropriate sensor features. Therefore, this paper proposes a denoise transformer Auto-Encoder (DTAE) as pre-processor for the tool condition monitoring (TCM) classifiers. The reconstruction task of the DTAE allows the model to pay more attention to the intrinsic wear features and the appropriate selection of them from the multi-sensor signals during feature extraction. Moreover, the loss function for DTAE ResNet is formed by summing the reconstruction loss from DTAE with the classification loss from ResNet. Compared DTAE ResNet to stacked sparse auto-encoder network, deep stacked auto-encoder network, ResNet-18, VGG-16, and LSTM, experiments demonstrated that the present method would attain the highest classification accuracies for tool wear suitable for TCM. And a comparative experiment was conducted to verify the effectiveness of the DTAE preprocessor in improving the anti-noise performance of the classifier.

Brain Image Segmentation for Ultrascale Neuron Reconstruction via an Adaptive Dual-Task Learning Network.

Accurate morphological reconstruction of neurons in whole brain images is critical for brain science research. However, due to the wide range of whole brain imaging, uneven staining, and optical system fluctuations, there are significant differences in image properties between different regions of the ultrascale brain image, such as dramatically varying voxel intensities and inhomogeneous distribution of background noise, posing an enormous challenge to neuron reconstruction from whole brain images. In this paper, we propose an adaptive dual-task learning network (ADTL-Net) to quickly and accurately extract neuronal structures from ultrascale brain images. Specifically, this framework includes an External Features Classifier (EFC) and a Parameter Adaptive Segmentation Decoder (PASD), which share the same Multi-Scale Feature Encoder (MSFE). MSFE introduces an attention module named Channel Space Fusion Module (CSFM) to extract structure and intensity distribution features of neurons at different scales for addressing the problem of anisotropy in 3D space. Then, EFC is designed to classify these feature maps based on external features, such as foreground intensity distributions and image smoothness, and select specific PASD parameters to decode them of different classes to obtain accurate segmentation results. PASD contains multiple sets of parameters trained by different representative complex signal-to-noise distribution image blocks to handle various images more robustly. Experimental results prove that compared with other advanced segmentation methods for neuron reconstruction, the proposed method achieves state-of-the-art results in the task of neuron reconstruction from ultrascale brain images, with an improvement of about 49% in speed and 12% in F1 score.

Application of a deep learning method for shower axis reconstruction in a 3D imaging calorimeter

We present a deep learning method based on Convolutional Neural Networks (CNN) to reconstruct the shower axis of isotropic electrons in a three-dimensional (3D) imaging calorimeter, which is one of the core instruments of the High Energy cosmic-Radiation Detection (HERD) payload. The CNN method is evaluated against the Principal Component Analysis (PCA) method, utilizing isotropic Monte Carlo (MC) simulation data covering an energy spectrum from 10 GeV to 1000 GeV, as well as beam test data ranging from 50 GeV to 250 GeV. The comparative analysis results indicate that the CNN outperforms the PCA by about 40% at 100 GeV in terms of angular resolution under isotropic conditions. Both methods are further validated using beam test data. The results demonstrate the effectiveness of the CNN method in reconstructing the shower axis, and they underscore the potential of incorporating advanced deep learning techniques into the comprehensive task of calorimeter reconstruction.

VNVC: A Versatile Neural Video Coding Framework for Efficient Human-Machine Vision.

Almost all digital videos are coded into compact representations before being transmitted. Such compact representations need to be decoded back to pixels before being displayed to humans and - as usual - before being enhanced/analyzed by machine vision algorithms. Intuitively, it is more efficient to enhance/analyze the coded representations directly without decoding them into pixels. Therefore, we propose a versatile neural video coding (VNVC) framework, which targets learning compact representations to support both reconstruction and direct enhancement/analysis, thereby being versatile for both human and machine vision. Our VNVC framework has a feature-based compression loop. In the loop, one frame is encoded into compact representations and decoded to an intermediate feature that is obtained before performing reconstruction. The intermediate feature can be used as reference in motion compensation and motion estimation through feature-based temporal context mining and cross-domain motion encoder-decoder to compress the following frames. The intermediate feature is directly fed into video reconstruction, video enhancement, and video analysis networks to evaluate its effectiveness. The evaluation shows that our framework with the intermediate feature achieves high compression efficiency for video reconstruction and satisfactory task performances with lower complexities.