Spatial Kernel Research Articles

Generalized Space-Time Autoregressive Modeling of the Vertical Distribution of Copper and Gold Grades with a Porphyry-Deposit Case Study

We examined the first-order application of the generalized space-time autoregressive GSTAR (1;1) model. The autoregressive model was used and was performed simultaneously in multiple drill-hole locations. The GSTAR model was applied to data with absolute time parameter units, such as hours, days, months, or years. Here a new perspective on modeling space-time data is raised. We used the relative time parameter index as a discretization of the same drilling depth of mineralization through a porphyritic deposit. Random variables were the copper and gold grades derived from the hydrothermal fluid that passed through the rock fractures in a porphyry copper deposit in Indonesia. This research aims to model the vertical distribution of copper and gold grades through backcasting the GSTAR (1;1) model. Such results could help geologists to predict copper and gold grades in deeper zones in an ore deposit. Two spatial weight matrices were used in the GSTAR (1;1) model, and these were based on a Euclidean distance and kernel function. Both weight matrices were constructed from different perspectives. The Euclidean distance approach gave a fixed weight matrix. Meanwhile, the kernel function approach gave the possibility to be random since it is based on real observations. It is obtained that the estimated (in-sample) and predicted (out-sample) kernel weight approach was accurate. Copper and gold grades data could recommend the GSTAR (1;1) model with a spatial kernel weight for modeling the vertical continuity case.

Geographically convolutional neural network weighted regression: a method for modeling spatially non-stationary relationships based on a global spatial proximity grid

Geographically weighted regression (GWR) is a classical method of modeling spatially non-stationary relationships. The geographically neural network weighted regression (GNNWR) model solves the problem of the inaccurate construction of spatial weight kernels using a spatially weighted neural network. However, when the spatial distribution of observations is uneven, the spatial proximity expression in the input of GWR and GNNWR models does not fully represent the impact of the whole research space on the estimating point. Therefore, we established a global spatial proximity grid (GSPG) to express the spatial proximity of each estimating point and proposed a spatially weighted convolutional neural network (SWCNN) to extract the relationship between the GSPG and spatial weights. Finally, we proposed a geographically convolutional neural network weighted regression (GCNNWR) model combining SWCNN and ordinary linear regression (OLR) model to estimate spatial non-stationarity. We used two case studies of simulated data and real environment data to demonstrate the advancements of the GCNNWR model. The GCNNWR model achieved higher estimation accuracy and greater predictive power than the OLR, GWR, multi-scale GWR (MGWR), and GNNWR models. Moreover, the GCNNWR model maintained its better stability and accuracy in estimating spatially non-stationary relationships when the distribution of observations was uneven.

Spatial characteristics and driving factors of urban flooding in Chinese megacities

Changes in hydrological processes caused by rapid urbanization lead to the growing incidence of urban flooding, which is a major challenge to urban sustainability. Urban floods have seriously threatened the natural environment and human life. Understanding the spatial patterns and influencing factors of urban flooding has important implications for mitigating urban flood hazards. Previous studies have demonstrated the impact of natural and human factors on urban flooding, but a comprehensive understanding of the mechanisms of urban flooding requires appropriate scales and evidence from multiple cities. In this study, 1201 flood records from 2015 to 2020 in 9 megacities (including Beijing, Tianjin, Shanghai, Xi’an, Nanjing, Wuhan, Guangzhou, Shenzhen, Shenyang) in China were used to investigate the spatial characteristics and driving factors of urban flooding. A combination of multiple stepwise regression and boosted regression tree (BRT) models was used to reveal the flood driving mechanism in megacities. The results indicated that flood events in 9 cities presented an aggregation effect, but the spatial characteristics and kernel density were different between coastal and inland cities. At all scales (1 km, 3 km, 5 km), physical geomorphic features (2D and 3D landscapes) have a high contribution to urban flooding, especially patch density, density of buildings and building shape coefficients. The relative contributions of the 2D and 3D landscape pattern factors increased by 43.8 % and 36.2 % as the grid scale increased from 1 km to 5 km, respectively. However, the relative contributions of topography, meteorology and drainage capacity factors decreased by 47.6 %, 39.0 % and 33.2 %, respectively. At a large scale (5 km), the correlation of driving factors for urban flooding was stronger, and the dominant factors were more obvious. At a small scale (1 km), the contribution of driving factors is relatively average. The scale effect significantly affects urban flooding. which suggests that an appropriate scale can more accurately capture the dominant drivers of urban flooding. Therefore, a novel method that integrates the stepwise regression model and BRT model were presented to quantify the complex relationship between urban flooding and driving factors under various analysis scales. The methods and results proposed by our study provided important insights and perspectives for urban flood risk mitigation and urban planning through multiscale analysis of flood driving mechanisms.

Study on the Spatial Pattern Characteristics and Influencing Factors of Inefficient Urban Land Use in the Yellow River Basin

In order to realize the optimization of urban spatial patterns in the Yellow River Basin, a study on the inefficient use of urban land in the Yellow River Basin was carried out. In this study, Dali County and Hancheng County in Weinan City are selected as the research areas. Firstly, the analytic hierarchy process is used to build a comprehensive evaluation system for the identification of inefficient land in stock; secondly, the standard deviation ellipse method and spatial kernel density estimation method are used to quantitatively analyze the spatial distribution characteristics of inefficient land. Thirdly, the contribution model is used to analyze the influencing factors of inefficient land use. Finally, corresponding redevelopment suggestions are given for each type of inefficient land. The results show that Dali had the smallest area of inefficient land; second is Xincheng Street in Hancheng City; and Longmen Town, Hancheng City has the largest area. The distribution of inefficient land in Dali and Longmen Town in Hancheng City is relatively balanced, while the distribution of all kinds of inefficient land in Xincheng Street in Hancheng City is not concentrated. The density of the road network is the most important contributing factor to inefficient land use in the study area. This paper comprehensively uses the methods of economics and geography to study inefficient land use, quantifies the spatial-temporal characteristics and influencing factors of land use units, explores the spatial patterns of land use and enriches the research into relevant theories.

The Temporal and Spatial Epidemiology Employed in the Elimination of the HIV Epidemic in the Largest Capital of the Brazilian Rainforest.

Background: The main goal of this study was to analyze the human immunodeficiency virus (HIV) epidemic temporally and spatially in Belém from 2007 to 2018. Methods: The incidence rates were analyzed according to time using autoregressive integrated moving-average models, as well as spatially using spatial autocorrelation, Kernel density, scan statistics, and regression techniques. Results: During the study period, 6007 notifications of new cases of HIV/AIDS were reported. The time series analysis revealed a stabilized trend of incidence from 2007 to October 2016, followed by irregular fluctuations until the end of December 2018. Seasonal behavior was observed from 2019 to 2022. The high–high incidence clusters were found in the central and transition areas. An expansion of the number of new reported cases was observed in the central area. Three spatial risk zones were observed. The higher relative risk zone was concentrated in the transition area. The spatial regression showed that the incidence rates were positively correlated with the Family Health Strategy (FHS) coverage. Conclusions: To eliminate HIV in Belém, it will be necessary to decentralize testing and ART and expand the coverage of FHS to ensure universal access to healthcare for citizens.

A Hyperspectral Image Classification Method Based on Adaptive Spectral Spatial Kernel Combined with Improved Vision Transformer

In recent years, methods based on deep convolutional neural networks (CNNs) have dominated the classification task of hyperspectral images. Although CNN-based HSI classification methods have the advantages of spatial feature extraction, HSI images are characterized by approximately continuous spectral information, usually containing hundreds of spectral bands. CNN cannot mine and represent the sequence properties of spectral features well, and the transformer model of attention mechanism proves its advantages in processing sequence data. This study proposes a new spectral spatial kernel combined with the improved Vision Transformer (ViT) to jointly extract spatial spectral features to complete classification task. First, the hyperspectral data are dimensionally reduced by PCA; then, the shallow features are extracted with an spectral spatial kernel, and the extracted features are input into the improved ViT model. The improved ViT introduces a re-attention mechanism and a local mechanism based on the original ViT. The re-attention mechanism can increase the diversity of attention maps at different levels. The local mechanism is introduced into ViT to make full use of the local and global information of the data to improve the classification accuracy. Finally, a multi-layer perceptron is used to obtain the classification result. Among them, the Focal Loss function is used to increase the loss weight of small-class samples and difficult-to-classify samples in HSI data samples and reduce the loss weight of easy-to-classify samples, so that the network can learn more useful hyperspectral image information. In addition, using the Apollo optimizer to train the HSI classification model to better update and compute network parameters that affect model training and model output, thereby minimizing the loss function. We evaluated the classification performance of the proposed method on four different datasets, and achieved good classification results on urban land object classification, crop classification and mineral classification, respectively. Compared with the state-of-the-art backbone network, the method achieves a significant improvement and achieves very good classification accuracy.

Quantum mechanics with spatial non-local effects and position-dependent mass

A new higher order Schrödinger equation characterized by a position-dependent mass is introduced based on long-range spatial kernel effects and von Roos arguments. The extended Schrödinger equation depends on the sign of the moments M k , k = 0 , 1 , 2 , … and a stabilized quantum dynamic is realized for M 2 > 0 and M 4 > 0 . We have discussed its implications in several quantum mechanical systems where more than a few were raised, mainly the emergence of the quantum Pais-Uhlenbeck and the relativistic quantum harmonic oscillators besides the complex periodic potential characterized by a PT symmetry.

The role of geographical spreaders in infectious pattern formation and front propagation speeds

The pattern formation and spatial spread of infectious populations are investigated using a kernel-based Susceptible–Infectious–Recovered (SIR) model applicable across a wide range of basic reproduction numbers Ro. The goal is to examine the role of geographical spreaders on transient spatial pattern formation of infectious populations and the associated maximum invasive front speeds cmax. In the simulations conducted here, geographical spreaders are defined as a portion of the infected population ϕ experiencing high mobility between identical communities. The spatial organization of the infected population and cmax are determined when the infections are randomly initiated in space within multiple communities. For small but finite ϕ, scaling analysis and numerical simulations in 1-dimension suggest that cmax∼γ(Ro−1) when the spreading kernel is Gaussian-shaped, where γ is the inverse of the infectious duration. This finding for cmax agrees with a diffusion-based representation of mobility in 1-D. Numerical simulations in 2-D across wide-ranging Ro suggest that σ, the variance of the spatial kernel describing mobility of long-distance geographical spreaders across communities, determines the spatial organization of infections across communities. When σ/Lr>5 (long-distance mobility, where Lr is the minimum spatial extent defining adjacent communities), the infectious population will experience a transient but spatially coherent pattern with a wavelength that can be derived from the spreading kernel properties. Moreover, the 2-D simulations for the bounded kernel suggest that attainment of cmax is also dictated by ϕ but the magnitude is not sensitive to ϕ unlike diffusion-based models.

Guided Gaussian range kernel filtering based on similarity-aware window

We present a window-adaptive Gaussian guided filtering method to smooth images while preserving edge features. The key to our algorithm is designing a similarity-aware filtering window based on density clustering to protect the edge structure and introducing a small-scale Gaussian spatial kernel as the input of the Gaussian range kernel to construct a guided filter for image smoothing. Specifically, we first utilized the Gaussian spatial kernel filtering with a small spatial bandwidth to yield a guidance input. Then, the Mahalanobis metric was used to calculate the distance between the center point and the other pixels in the box filtering window, by which we employed the center density clustering algorithm to obtain a better nonbox region (i.e., similarity-aware window) in which each pixel was similar to the center point. Finally, based on the guidance input and the similarity-aware window, the guided Gaussian range filter performed better on image smoothing. Our proposed algorithm is simple and easy to implement. In particular, the window-aware technology effectively improved edge protection, and it can be widely used in image denoising, background smoothing, detexturing, detail enhancement, and edge extraction.

Data fusion reconstruction of spatially embedded complex networks

Abstract We introduce a kernel Lasso (kLasso) approach which is a type of sparse optimization that simultaneously accounts for spatial regularity and structural sparsity to reconstruct spatially embedded complex networks from time-series data about nodal states. Through the design of a spatial kernel function motivated by real-world network features, the proposed kLasso approach exploits spatial embedding distances to penalize overabundance of spatially long-distance connections. Examples of both random geometric graphs and real-world transportation networks show that the proposed method improves significantly upon existing network reconstruction techniques that mainly concern sparsity but not spatial regularity. Our results highlight the promise of data and information fusion in the reconstruction of complex networks, by utilizing both microscopic node-level dynamics (e.g. time series data) and macroscopic network-level information (metadata or other prior information).

Bi-volution: A Static and Dynamic Coupled Filter

Dynamic convolution has achieved significant gain in performance and computational complexity, thanks to its powerful representation capability given limited filter number/layers. However, SOTA dynamic convolution operators are sensitive to input noises (e.g., Gaussian noise, shot noise, e.t.c.) and lack sufficient spatial contextual information in filter generation. To alleviate this inherent weakness, we propose a lightweight and heterogeneous-structure (i.e., static and dynamic) operator, named Bi-volution. On the one hand, Bi-volution is designed as a dual-branch structure to fully leverage complementary properties of static/dynamic convolution, which endows Bi-volution more robust properties and higher performance. On the other hand, the Spatial Augmented Kernel Generation module is proposed to improve the dynamic convolution, realizing the learning of spatial context information with negligible additional computational complexity. Extensive experiments illustrate that the ResNet-50 equipped with Bi-volution achieves a highly competitive boost in performance (+2.8% top-1 accuracy on ImageNet classification, +2.4% box AP and +2.2% mask AP on COCO detection and instance segmentation) while maintaining extremely low FLOPs (i.e., ResNet50@2.7 GFLOPs). Furthermore, our Bi-volution shows better robustness than dynamic convolution against various noise and input corruptions. Our code is available at https://github.com/neuralchen/Bivolution.

Adaptive Bilateral Texture Filter for Image Smoothing.

The biggest challenge of texture filtering is to smooth the strong gradient textures while maintaining the weak structures, which is difficult to achieve with current methods. Based on this, we propose a scale-adaptive texture filtering algorithm in this paper. First, the four-directional detection with gradient information is proposed for structure measurement. Second, the spatial kernel scale for each pixel is obtained based on the structure information; the larger spatial kernel is for pixels in textural regions to enhance the smoothness, while the smaller spatial kernel is for pixels on structures to maintain the edges. Finally, we adopt the Fourier approximation of range kernel, which reduces computational complexity without compromising the filtering visual quality. By subjective and objective analysis, our method outperforms the previous methods in eliminating the textures while preserving main structures and also has advantages in structure similarity and visual perception quality.

Detection of Standing Dead Trees after Pine Wilt Disease Outbreak with Airborne Remote Sensing Imagery by Multi-Scale Spatial Attention Deep Learning and Gaussian Kernel Approach

The continuous and extensive pinewood nematode disease has seriously threatened the sustainable development of forestry in China. At present, many studies have used high-resolution remote sensing images combined with a deep semantic segmentation algorithm to identify standing dead trees in the red attack period. However, due to the complex background, closely distributed detection scenes, and unbalanced training samples, it is difficult to detect standing dead trees (SDTs) in a variety of complex scenes by using conventional segmentation models. In order to further solve the above problems and improve the recognition accuracy, we proposed a new detection method called multi-scale spatial supervision convolutional network (MSSCN) to identify SDTs in a wide range of complex scenes based on airborne remote sensing imagery. In the method, a Gaussian kernel approach was used to generate a confidence map from SDTs marked as points for training samples, and a multi-scale spatial attention block was added into fully convolutional neural networks to reduce the loss of spatial information. Further, an augmentation strategy called copy–pasting was used to overcome the lack of efficient samples in this research area. Validation at four different forest areas belonging to two forest types and two diseased outbreak intensities showed that (1) the copy–pasting method helps to augment training samples and can improve the detecting accuracy with a suitable oversampling rate, and the best oversampling rate should be carefully determined by the input training samples and image data. (2) Based on the two-dimensional spatial Gaussian kernel distribution function and the multi-scale spatial attention structure, the MSSCN model can effectively find the dead tree extent in a confidence map, and by following this with maximum location searching we can easily locate the individual dead trees. The averaged precision, recall, and F1-score across different forest types and disease-outbreak-intensity areas can achieve 0.94, 0.84, and 0.89, respectively, which is the best performance among FCN8s and U-Net. (3) In terms of forest type and outbreak intensity, the MSSCN performs best in pure pine forest type and low-outbreak-intensity areas. Compared with FCN8s and U-Net, the MSSCN can achieve the best recall accuracy in all forest types and outbreak-intensity areas. Meanwhile, the precision metric is also maintained at a high level, which means that the proposed method provides a trade-off between the precision and recall in detection accuracy.

Multi-scale ResNet and BiGRU automatic sleep staging based on attention mechanism.

Sleep staging is the basis of sleep evaluation and a key step in the diagnosis of sleep-related diseases. Despite being useful, the existing sleep staging methods have several disadvantages, such as relying on artificial feature extraction, failing to recognize temporal sequence patterns in the long-term associated data, and reaching the accuracy upper limit of sleep staging. Hence, this paper proposes an automatic Electroencephalogram (EEG) sleep signal staging model, which based on Multi-scale Attention Residual Nets (MAResnet) and Bidirectional Gated Recurrent Unit (BiGRU). The proposed model is based on the residual neural network in deep learning. Compared with the traditional residual learning module, the proposed model additionally uses the improved channel and spatial feature attention units and convolution kernels of different sizes in parallel at the same position. Thus, multiscale feature extraction of the EEG sleep signals and residual learning of the neural networks is performed to avoid network degradation. Finally, BiGRU is used to determine the dependence between the sleep stages and to realize the automatic learning of sleep data staging features and sleep cycle extraction. According to the experiment, the classification accuracy and kappa coefficient of the proposed method on sleep-EDF data set are 84.24% and 0.78, which are respectively 0.24% and 0.21 higher than the traditional residual net. At the same time, this paper also verified the proposed method on UCD and SHHS data sets, and the figure of classification accuracy is 79.34% and 81.6%, respectively. Compared to related existing studies, the recognition accuracy is significantly improved, which validates the effectiveness and generalization performance of the proposed method.

Point-of-interest recommendation in location-based social networks based on collaborative filtering and spatial kernel weighting

The growing popularity of location-based social networks (LBSNs), as well as their financial benefits, has led to a variety of methods for point-of-interest (POI) recommendations. A review of various methodologies reveals the role of geography is rarely addressed individually in current procedures. Instead, it is treated as an overall probability distribution of distances between all check-in pairs. On the other hand, despite the increased accuracy, the incorporation of individualized probabilistic distributions complicates matters of recommender systems. To address this issue, the main objective of this study is to introduce a POI recommendation mechanism in LBSNs that could be simple to implement the geographical influence. Furthermore, the method could be easily integrated with collaborative filtering while also increasing the quality of recommendations. Hence, instead of establishing a unique probability function for all check-ins data using the power-law or the Gaussian distribution, a spatial kernel weighting technique inspired by the similarity-weighted average notion, called CFSKW, was introduced and implemented. CFSKW was found on the principle that POIs close to a person's favourite POIs are more likely to be visited. As expected, the proposed method outperformed state-of-the-arts when tested utilizing two Four square datasets, namely New York and Tokyo. Kernel bandwidth and the coefficient of integrating user preference with geographical influence were two critical factors to be tuned. Low bandwidth and a medium to high geographical coefficient, according to the findings, produce better results. In unproportionate density of POIs, like that of the Tokyo dataset, a proposed algorithm based on dynamic bandwidth showed promising results. For the New York dataset, Precision@5 and Recall@5 were obtained as 25.41% and 61.80% using the proposed method, respectively, indicating a 3% and 5.1% improvement over the reference techniques. Precision@5 and Recall@5 were 1.1% and 1.2% better for the Tokyo dataset, respectively, when using a dynamic bandwidth geographical similarity algorithm.

Multiframe Correction Blind Deconvolution for Solar Image Restoration

A series of short-exposure images are often used for rich, small-scale structure, high-quality, and high-resolution astronomical observations. Postprocessing of the closed-loop adaptive optics (AO) image using ground-based astronomical telescopes plays an important role in astronomical observations due to it further improving image quality after AO processing. These images show several main characteristics: random spatial variation blur kernel, unclear model after AO correction, unclear physical characteristics of observation objects, etc. Our goal is to propose a multiframe correction blind deconvolution (MFCBD) algorithm to restore AO closed-loop solar images. MFCBD introduces a denoiser and corrector to help estimate the intermediate latent image and proposes using an L q norm of the kernel as the sparse constraint to acquire a compact blur kernel. MFCBD also uses the half-quadratic splitting strategy to optimize the objective function, which makes the algorithm not only simple to solve, but also easy to adapt to different fidelity terms and prior terms. In tests on three data sets observed from the photosphere and chromosphere of the Sun, MFCBD not only restored clearer and more detailed images, but also converged smoothly and monotonically in terms of the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) after a few iterations. Taking the speckle-reconstructed image as a reference, the clear image restored by our method performs best both in PSNR and SSIM compared with the state-of-the-art traditional methods OBD and BATUD.

A Dynamic Convolution Kernel Generation Method Based on Regularized Pattern for Image Super-Resolution.

Image super-resolution aims to reconstruct a high-resolution image from its low-resolution counterparts. Conventional image super-resolution approaches share the same spatial convolution kernel for the whole image in the upscaling modules, which neglect the specificity of content information in different positions of the image. In view of this, this paper proposes a regularized pattern method to represent spatially variant structural features in an image and further exploits a dynamic convolution kernel generation method to match the regularized pattern and improve image reconstruction performance. To be more specific, first, the proposed approach extracts features from low-resolution images using a self-organizing feature mapping network to construct regularized patterns (RP), which describe different contents at different locations. Second, the meta-learning mechanism based on the regularized pattern predicts the weights of the convolution kernels that match the regularized pattern for each different location; therefore, it generates different upscaling functions for images with different content. Extensive experiments are conducted using the benchmark datasets Set5, Set14, B100, Urban100, and Manga109 to demonstrate that the proposed approach outperforms the state-of-the-art super-resolution approaches in terms of both PSNR and SSIM performance.

Modelling the emergence of cities andurban patterning using coupled integro-differential equations.

Human residential population distributions show patterns of higher density clustering around local services such as shops and places of employment, displaying characteristic length scales; Fourier transforms and spatial autocorrelation show the length scale between UK cities is around 45 km. We use integro-differential equations to model the spatio-temporal dynamics of population and service density under the assumption that they benefit from spatial proximity, captured via spatial weight kernels. The system tends towards a well-mixed homogeneous state or a spatial pattern. Linear stability analysis around the homogeneous steady state predicts a modelled length-scale consistent with that observed in the data. Moreover, we show that spatial instability occurs only for perturbations with a sufficiently long wavelength and only where there is a sufficiently strong dependence of service potential on population density. Within urban centres, competition for space may cause services and population to be out of phase with one another, occupying separate parcels of land. By introducing competition, along with a preference for population to be located near, but not too near, to high service density areas, secondary out-of-phase patterns occur within the model, at a higher density and with a shorter length scale than in phase patterning. Thus, we show that a small set of core behavioural ingredients can generate aggregations of populations and services, and pattern formation within cities, with length scales consistent with real-world data. The analysis and results are valid across a wide range of parameter values and functional forms in the model.

Solving three major biases of the ETAS model to improve forecasts of the 2019 Ridgecrest sequence

Strong earthquakes cause aftershock sequences that are clustered in time according to a power decay law, and in space along their extended rupture, shaping a typically elongate pattern of aftershock locations. A widely used approach to model earthquake clustering, the Epidemic Type Aftershock Sequence (ETAS) model, shows three major biases. First, the conventional ETAS approach assumes isotropic spatial triggering, which stands in conflict with observations and geophysical arguments for strong earthquakes. Second, the spatial kernel has unlimited extent, allowing smaller events to exert disproportionate trigger potential over an unrealistically large area. Third, the ETAS model assumes complete event records and neglects inevitable short-term aftershock incompleteness as a consequence of overlapping coda waves. These three aspects can substantially bias the parameter estimation and lead to underestimated cluster sizes. In this article, we combine the approach of Grimm et al. (Bulletin of the Seismological Society of America, 2021), who introduced a generalized anisotropic and locally restricted spatial kernel, with the ETAS-Incomplete (ETASI) time model of Hainzl (Bulletin of the Seismological Society of America, 2021), to define an ETASI space-time model with flexible spatial kernel that solves the abovementioned shortcomings. We apply different model versions to a triad of forecasting experiments of the 2019 Ridgecrest sequence, and evaluate the prediction quality with respect to cluster size, largest aftershock magnitude and spatial distribution. The new model provides the potential of more realistic simulations of on-going aftershock activity, e.g. allowing better predictions of the probability and location of a strong, damaging aftershock, which might be beneficial for short term risk assessment and disaster response.

A Principal Neighborhood Aggregation-Based Graph Convolutional Network for Pneumonia Detection.

Pneumonia is one of the main causes of child mortality in the world and has been reported by the World Health Organization (WHO) to be the cause of one-third of child deaths in India. Designing an automated classification system to detect pneumonia has become a worthwhile research topic. Numerous deep learning models have attempted to detect pneumonia by applying convolutional neural networks (CNNs) to X-ray radiographs, as they are essentially images and have achieved great performances. However, they failed to capture higher-order feature information of all objects based on the X-ray images because the topology of the X-ray images’ dimensions does not always come with some spatially regular locality properties, which makes defining a spatial kernel filter in X-ray images non-trivial. This paper proposes a principal neighborhood aggregation-based graph convolutional network (PNA-GCN) for pneumonia detection. In PNA-GCN, we propose a new graph-based feature construction utilizing the transfer learning technique to extract features and then construct the graph from images. Then, we propose a graph convolutional network with principal neighborhood aggregation. We integrate multiple aggregation functions in a single layer with degree-scalers to capture more effective information in a single layer to exploit the underlying properties of the graph structure. The experimental results show that PNA-GCN can perform best in the pneumonia detection task on a real-world dataset against the state-of-the-art baseline methods.