Earth Mover's Distance Research Articles

A coarse-to-fine point completion network with details compensation and structure enhancement

Point cloud completion, the issue of estimating the complete geometry of objects from partially-scanned point cloud data, becomes a fundamental task in many 3d vision and robotics applications. To address the limitations on inadequate prediction of shape details for traditional methods, a novel coarse-to-fine point completion network (DCSE-PCN) is introduced in this work using the modules of local details compensation and shape structure enhancement for effective geometric learning. The coarse completion stage of our network consists of two branches—a shape structure recovery branch and a local details compensation branch, which can recover the overall shape of the underlying model and the shape details of incomplete point cloud through feature learning and hierarchical feature fusion. The fine completion stage of our network employs the structure enhancement module to reinforce the correlated shape structures of the coarse repaired shape (such as regular arrangement or symmetry), thus obtaining the completed geometric shape with finer-grained details. Extensive experimental results on ShapeNet dataset and ModelNet dataset validate the effectiveness of our completion network, which can recover the shape details of the underlying point cloud whilst maintaining its overall shape. Compared to the existing methods, our coarse-to-fine completion network has shown its competitive performance on both chamfer distance (CD) and earth mover distance (EMD) errors. Such as, the repairing results on the ShapeNet dataset of our completion network are reduced by an average of 35.62%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$35.62\\%$$\\end{document}, 33.31%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$33.31\\%$$\\end{document}, 29.62%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$29.62\\%$$\\end{document}, and 23.62%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$23.62\\%$$\\end{document} in terms of CD error, comparing with PCN, FoldingNet, Atlas, and CRN methods, respectively; and also reduced by an average of 15.63%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$15.63\\%$$\\end{document}, 1.29%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$1.29\\%$$\\end{document}, 64.52%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$64.52\\%$$\\end{document}, and 62.87%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$62.87\\%$$\\end{document} in terms of EMD error, respectively. Meanwhile, the completion results on the ModelNet dataset of our network have an average reduction of 28.41%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$28.41\\%$$\\end{document}, 26.57%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$26.57\\%$$\\end{document}, 20.65%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$20.65\\%$$\\end{document}, and 18.55%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$18.55\\%$$\\end{document} in terms of CD error, comparing to PCN, FoldingNet, Atlas, and CRN methods, respectively; and also an average reduction of 21.91%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$21.91\\%$$\\end{document}, 19.59%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$19.59\\%$$\\end{document}, 43.51%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$43.51\\%$$\\end{document}, and 21.49%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$21.49\\%$$\\end{document} in terms of EMD error, respectively. Our proposed point completion network is also robust to different degrees of data incompleteness and model noise.

Design of 3D Environment Combining Digital Image Processing Technology and Convolutional Neural Network

As virtual reality technology advances, 3D environment design and modeling have garnered increasing attention. Applications in networked virtual environments span urban planning, industrial design, and manufacturing, among other fields. However, existing 3D modeling methods exhibit high reconstruction error precision, limiting their practicality in many domains, particularly environmental design. To enhance 3D reconstruction accuracy, this study proposes a digital image processing technology that combines binocular camera calibration, stereo correction, and a convolutional neural network (CNN) algorithm for optimization and improvement. By employing the refined stereo-matching algorithm, a 3D reconstruction model was developed to augment 3D environment design and reconstruction accuracy while optimizing the 3D reconstruction effect. An experiment using the ShapeNet dataset demonstrated that the evaluation indices—Chamfer distance (CD), Earth mover’s distance (EMD), and intersection over union—of the model constructed in this study outperformed those of alternative methods. After incorporating the CNN module in the ablation experiment, CD and EMD increased by an average of 0.1 and 0.06, respectively. This validates that the proposed CNN module effectively enhances point cloud reconstruction accuracy. Upon adding the CNN module, the CD index and EMD index in the dataset increased by an average of 0.34 and 0.54, respectively. These results indicate that the proposed CNN module exhibits strong predictive capabilities for point cloud coordinates. Furthermore, the model demonstrates good generalization performance.

Beyond correlation: optimal transport metrics for characterizing representational stability and remapping in neurons encoding spatial memory.

Spatial representations in the entorhinal cortex (EC) and hippocampus (HPC) are fundamental to cognitive functions like navigation and memory. These representations, embodied in spatial field maps, dynamically remap in response to environmental changes. However, current methods, such as Pearson's correlation coefficient, struggle to capture the complexity of these remapping events, especially when fields do not overlap, or transformations are non-linear. This limitation hinders our understanding and quantification of remapping, a key aspect of spatial memory function. We propose a family of metrics based on the Earth Mover's Distance (EMD) as a versatile framework for characterizing remapping. The EMD provides a granular, noise-resistant, and rate-robust description of remapping. This approach enables the identification of specific cell types and the characterization of remapping in various scenarios, including disease models. Furthermore, the EMD's properties can be manipulated to identify spatially tuned cell types and to explore remapping as it relates to alternate information forms such as spatiotemporal coding. We present a feasible, lightweight approach that complements traditional methods. Our findings underscore the potential of the EMD as a powerful tool for enhancing our understanding of remapping in the brain and its implications for spatial navigation, memory studies and beyond.

Calculating Pairwise Similarity of Polymer Ensembles via Earth Mover's Distance.

Synthetic polymers, in contrast to small molecules and deterministic biomacromolecules, are typically ensembles composed of polymer chains with varying numbers, lengths, sequences, chemistry, and topologies. While numerous approaches exist for measuring pairwise similarity among small molecules and sequence-defined biomacromolecules, accurately determining the pairwise similarity between two polymer ensembles remains challenging. This work proposes the earth mover's distance (EMD) metric to calculate the pairwise similarity score between two polymer ensembles. EMD offers a greater resolution of chemical differences between polymer ensembles than the averaging method and provides a quantitative numeric value representing the pairwise similarity between polymer ensembles in alignment with chemical intuition. The EMD approach for assessing polymer similarity enhances the development of accurate chemical search algorithms within polymer databases and can improve machine learning techniques for polymer design, optimization, and property prediction.

Matching Distributions Algorithms Based on the Earth Mover's Distance for Ordinal Quantification.

The goal of quantification learning is to induce models capable of accurately predicting the class distribution for new bags of unseen examples. These models only return the prevalence of each class in the bag because prediction of individual examples is irrelevant in these tasks. A prototypical application of ordinal quantification is to predict the proportion of opinions that fall into each category from one to five stars. Ordinal quantification has hardly been studied in the literature, and in fact, only one approach has been proposed so far. This article presents a comprehensive study of ordinal quantification, analyzing the applicability of the most important algorithms devised for multiclass quantification and proposing three new methods that are based on matching distributions using Earth mover's distance (EMD). Empirical experiments compare 14 algorithms on synthetic and benchmark data. To statistically analyze the obtained results, we further introduce an EMD-based scoring function. The main conclusion is that methods using a criterion somehow related to EMD, including two of our proposals, obtain significantly better results.

Research on incentive mechanisms for anti-heterogeneous federated learning based on reputation and contribution

<abstract> <p>An optimization algorithm for federated learning, equipped with an incentive mechanism, is introduced to tackle the challenges of excessive iterations, prolonged training durations, and suboptimal efficiency encountered during model training within the federated learning framework. Initially, the algorithm establishes reputation values that are tied to both time and model loss metrics. This foundation enables the creation of incentive mechanisms aimed at rewarding honest nodes while penalizing malicious ones. Subsequently, a bidirectional selection mechanism anchored in blockchain technology is developed, allowing smart contracts to enroll nodes with high reputations in training sessions, thus filtering out malicious clients and enhancing local training efficiency. Furthermore, the integration of the Earth Mover's Distance (EMD) mechanism serves to lessen the impact of non-IID (non-Independent and Identically Distributed) data on the global model, leading to a reduction in the frequency of model training cycles and an improvement in model accuracy. Experimental results confirm that this approach maintains high model accuracy in non-IID data settings, outperforming traditional federated learning algorithms.</p> </abstract>

Testing Distributions of Huge Objects

We initiate a study of a new model of property testing that is a hybrid of testing properties of distributions and testing properties of strings. Specifically, the new model refers to testing properties of distributions, but these are distributions over huge objects (i.e., very long strings). Accordingly, the model accounts for the total number of local probes into these objects (resp., queries to the strings) as well as for the distance between objects (resp., strings), and the distance between distributions is defined as the earth mover's distance with respect to the relative Hamming distance between strings. We study the query complexity of testing in this new model, focusing on three directions. First, we try to relate the query complexity of testing properties in the new model to the sample complexity of testing these properties in the standard distribution testing model. Second, we consider the complexity of testing properties that arise naturally in the new model (e.g., distributions that capture random variations of fixed strings). Third, we consider the complexity of testing properties that were extensively studied in the standard distribution testing model: Two such cases are uniform distributions and pairs of identical distributions.

Efficient Approximation Framework for Attribute Recommendation

Trend analysis is a fundamental type of analytical query in online analytical processing (OLAP) systems. In trend analysis, a key step is to identify k valuable attributes whose distributions in two subsets under different predicates significantly differ for further investigation, where the difference is measured by metric functions. However, the exact solution that involves scanning all records is prohibitively expensive, particularly when handling large datasets in the era of big data. To minimize unnecessary data access, the existing state-of-the-art solution TopKAttr adopts sampling to avoid the expensive data scan. However, their solution still has two main drawbacks. Firstly, their solution is tailored only for two limited metric functions: the Earth Mover distance and Euclidean distance, and cannot be generalized to more complicated metric functions. Besides, their solution still aims to return the exact top-k answers via the sampling method, which still causes high running costs as shown in our experiment. Motivated by these limitations, we propose a general approximation framework for attribute recommendation that efficiently returns the top-k attributes with theoretical guarantees while supporting an extensive range of metric functions, such as the Kolmogorov-Smirnov test (KS-test), Chebyshev distance, the Earth Mover distance, Euclidean distance, and with the potential to more metrics. The key to our framework is a new bound estimation strategy that can be applied to a wide spectrum of metrics, as we listed above. Based on our estimation framework, we further devise an efficient approximation algorithm with theoretical guarantees to answer the top-k queries, which is widely used in attribute recommendation. Extensive experiments on four real large datasets show that our framework gains up to an order of magnitude speed-up and consistently high accuracy compared to TopKAttr, providing a promising alternative for attribute recommendation in OLAP systems.

Differentiable Earth mover’s distance for data compression at the high-luminosity LHC

The Earth mover’s distance (EMD) is a useful metric for image recognition and classification, but its usual implementations are not differentiable or too slow to be used as a loss function for training other algorithms via gradient descent. In this paper, we train a convolutional neural network (CNN) to learn a differentiable, fast approximation of the EMD and demonstrate that it can be used as a substitute for computing-intensive EMD implementations. We apply this differentiable approximation in the training of an autoencoder-inspired neural network (encoder NN) for data compression at the high-luminosity LHC at CERN The goal of this encoder NN is to compress the data while preserving the information related to the distribution of energy deposits in particle detectors. We demonstrate that the performance of our encoder NN trained using the differentiable EMD CNN surpasses that of training with loss functions based on mean squared error.

A novel trajectory similarity measurement method based on node‐sequence hierarchical digraph

AbstractTrajectory similarity measurement is a basic and vital task in trajectory data mining, which has attracted extensive research in the past decades. Recent works focused on the sequence and hierarchy property of trajectories to construct similarity measurements. However, these methods ignore the user information on the visiting locations, such as semantic and time distribution. In light of this, a novel trajectory similarity measurement based on Node‐Sequence Hierarchical Digraph (NSHD) framework is proposed in this article. We first propose a Time‐Weighted Stay Point Detection (TWSPD) method to extract real visiting locations of users more accurately. Then, the nodes of digraph are obtained by clustering users' stay points and the edges of digraph are sequence information that users move between these nodes. An Advanced Earth Mover's Distance (AEMD) is proposed to measure the node similarity between users, considering visiting time distribution and semantic information simultaneously. Both node and sequence similarities are used to calculate the similarity score to obtain the final trajectory similarity measurement. Experiments on Geolife and T‐Drive datasets show that our proposed method offers competitive performance with mean reciprocal rank values reaching 96.01 and 81.26%, which outperforms related trajectory similarity measurements by more than 10 and 15%.

CWEmd: A Lightweight Similarity Measurement for Resource-Constrained Vehicular Networks

Generating an accurate machine learning (ML) model is of great importance for the Internet of Vehicles (IoV). However, obtaining such a model is challenging due to the fact that sub-groups of in-network vehicles receive data from different resources. A worthwhile investment then would be identifying those groups before inferring models. Similarity metrics are widely used to distinguish different groups. However, the efficiency of most existing similarity measurements is at the cost of increased computational complexity and decreased accuracy, making them unsuitable for IoV’s stringent conditions. To address this issue, we propose a computationally efficient method to measure the similarity of different vehicles, where a simplified version of Earth Mover’s Distance (EMD) is adopted. This distance metric is then embedded into a distributed clustering algorithm to learn the global pattern for vehicular systems. Our algorithm’s overall performance is measured using an Asynchronous Message Delay Simulator. Compared to the best algorithm of the state-of-the-art, our proposed algorithm converges slightly slower (by less than 1%) but improves the clustering accuracy by as much as 20% with synthetic data. Additionally, real-world data collected from Vehicles validates the efficiency of our proposed algorithm.

NIMG-84. COMPARISON OF TRADITIONAL AND ADVANCED DEEP-LEARNING PATIENT SPECIFIC REGISTRATION METHODS IN PRESERVING TUMOR HABITAT IN GLIOBLASTOMA TUMORS: A PRELIMINARY STUDY

Abstract PURPOSE Registration of patient specific pre and post operative MRI scans in Glioblastoma patients is crucial to assess tumor progression, effect of radiation, and chemotherapy and survival prediction. However, in brain tumors longitudinal registration is challenging as tumor resection and mass effects causes tissue deformations and missing intensity correspondence across longitudinal scans. In this study, we compared effectiveness of traditional registration algorithms with state-of-the-art computationally extensive deep learning (DL) models in preserving intensity of tumor regions. METHODS Brain MRI scans (T1w,T1c,T2w,FLAIR) of 14 glioblastoma patients with first (tp1), intermediate (tp2) and follow up post operative scans (tp3) were obtained from two publicly available dataset (LUMIERE, XCURE). Preprocessing involved co-registration of all modalities within a timepoint, skull stripping, and uniformity in image dimensions, orientation, and origin across all timepoints. Tumor scans at tp2 and tp3 were registered to tp1 for all modalities independently, using different registration algorithms such as the traditional affine and symmetric diffeomorphic (SyN) registration with and without tumor masks using the ANTs software, along with the MICCAI-registration challenge winner DL algorithm (DIRAC). Intensity distribution of registered and unregistered scans for tp2 and tp3 were compared using the earth mover’s distance (EMD). RESULTS Traditional affine (tp2-0.074±0.06, tp3:0.058±0.051) and SyN (tp2-0.048±0.031 ,tp3:0.075±0.074) registration algorithm using T2w modality provided lower average EMD than DL method (tp2-0.624±0.69,tp3-0.595±0.67). Intensity distribution was better preserved for scans from intermediate timepoints compared to the post operative scans. CONCLUSION Our preliminary assessment suggests that traditional registration methods when implemented on pre-aligned MRI scans may better preserve image intensities in brain tumors scans compared to advanced and computationally expensive DL methods.

Intelligent phenotype-detection and gene expression profile generation with generative adversarial networks

Gene expression analysis is valuable for cancer type classification and identifying diverse cancer phenotypes. The latest high-throughput RNA sequencing devices have enabled access to large volumes of gene expression data. However, we face several challenges, such as data security and privacy, when we develop machine learning-based classifiers for categorizing cancer types with these datasets. To address these issues, we propose IP3G (Intelligent Phenotype-detection and Gene expression profile Generation with Generative adversarial network), a model based on Generative Adversarial Networks. IP3G tackles two major problems: augmenting gene expression data and unsupervised phenotype discovery. By converting gene expression profiles into 2-Dimensional images and leveraging IP3G, we generate new profiles for specific phenotypes. IP3G learns disentangled representations of gene expression patterns and identifies phenotypes without labeled data. We improve the objective function of the GAN used in IP3G by employing the earth mover distance and a novel mutual information function. IP3G outperforms clustering methods like k-Means, DBSCAN, and GMM in unsupervised phenotype discovery, while also surpassing SVM and CNN classification accuracy by up to 6% through gene expression profile augmentation. The source code for the developed IP3G is accessible to the public on GitHub.

The study for optimization strategies on the performance of DCGAN

Since Deep Convolutional Generative Adversarial Networks (DCGAN) was proposed, it has been perceived as a model with difficulty in training due to several factors. To solve this problem, dozens of optimization strategies were presented, but none of them was compared with the others. In this paper, the author chose three representative methods, namely one-label smoothing, the two Time-Scale Update Rule (TTUR), and the Earth-Mover Distance (EMD) or Wasserstein-1 to make a comparison of the optimization effect on the DCGAN model. To be specific, these three approaches were adopted respectively while using MNIST and Fashion-MNIST as datasets. One-side label smoothing was designed to prevent overconfidence in the model by adding a penalty term in the discriminator. TTUR was a simpler update strategy that could help the model find the stationary local Nash equilibrium under mild assumptions. EMD was an alternative loss function that enabled the model to distinguish the difference while the real distribution and generated distribution were not overlapped. Contrast experiments were conducted both vertically and horizontally. The author applied these three methods with the same dataset and the same method with different datasets in order to compare the time of the model collapse, the trend of loss in line graphs, and the impact of different datasets on results. Experimental results indicated that both one-label smoothing and TTUR postponed the model collapse while EMD completely get rid of it. Furthermore, generated images may lose texture information when using more complicated datasets.

Deep learning-based software bug classification

Context:Accurate classification of bugs can help accelerate the bug triage process, code inspection, and repair activities. In this context, many machine learning techniques have been proposed to classify bugs. The expressive power of deep learning could be used to further improve classification. Objective:We propose a novel deep learning-based bug classification approach. Methods:We first build a bug taxonomy with eight bug classes, each characterized by a set of keywords. Subsequently, we heuristically annotate a moderately large set (∼1.36M) of software bug resolution reports using an earth-mover distance technique based on the keywords. Finally, we use four attention-based classification techniques to classify these curated bugs. Results:Our experiments on a carefully collected dataset indicate that our proposed technique achieved a mean F1-Score of 84.78% and a mean macro-average ROC of 98.25%. Conclusion:Our proposed approach was observed to outperform the existing techniques by 16.88% on an average in terms of F1-Score for the considered dataset.

Distance measures for geometric graphs

A geometric graph is a combinatorial graph, endowed with a geometry that is inherited from its embedding in a Euclidean space. Formulation of a meaningful measure of (dis-)similarity in both the combinatorial and geometric structures of two such geometric graphs is a challenging problem in pattern recognition. We study two notions of distance measures for geometric graphs, called the geometric edit distance (GED) and geometric graph distance (GGD). While the former is based on the idea of editing one graph to transform it into the other graph, the latter is inspired by inexact matching of the graphs. For decades, both notions have been lending themselves well as measures of similarity between attributed graphs. If used without any modification, however, they fail to provide a meaningful distance measure for geometric graphs—even cease to be a metric. We have curated their associated cost functions for the context of geometric graphs. Alongside studying the metric properties of GED and GGD, we investigate how the two notions compare. We further our understanding of the computational aspects of GGD by showing that the distance is NP-hard to compute, even if the graphs are planar and arbitrary cost coefficients are allowed.As a computationally tractable alternative, we propose in this paper the Graph Mover's Distance (GMD), which has been formulated as an instance of the earth mover's distance. The computation of the GMD between two geometric graphs with at most n vertices takes only O(n3)-time. The GMD demonstrates extremely promising empirical evidence at recognizing letter drawings.

AI Beam Angle Prediction in Proton Therapy for Brain Patients.

Patients receiving brain Intensity-modulated Proton Therapy (IMPT) are treated with 1-3 fields. Typically, beam angles must be selected manually, requiring substantial planner time and experience. The choice of beam angles has a major impact on the treatment plan quality. An AI model was developed to automate this process aiming to improve efficiency and potentially plan quality. AI performance was tested by comparing predicted beam angles to the human ground truth angles and evaluating plans resulting from each set of beam angles. Previously treated IMPT brain patients were divided into a training (n = 40) and validation set (n = 10) for the AI. Beam-angle selection in the AI was cast as a multi-label classification problem, training the convolution neural network with a Circular Earth Mover's Distance based regularization and multi-label circular-smooth label technique. An analytical post-processing algorithm is employed to minimize distance to the target and avoid organs-at-risk. Finalized gantry angle predictions are discretized in steps of 5°. For an independent test set of n = 10 patients beam angles were predicted by the AI and compared to those chosen by human planners. Both sets of angles were used to create treatment plans with an automated knowledge-based planning (KBP) tool for brain IMPT utilizing single-field optimization and robust optimization. Plan differences are therefore solely due to the choice of beam angles as automated KBP removes human optimization variability. Resulting plan quality was compared by standard clinical dosimetric parameters to the CTV, brain, brainstem, cochlea, cord, eye, lens, optic chiasm, optic nerves, pituitary, and temporal lobes. All Human and AI selected beam angles are shown in the table. For cases 1-6, AI and Human selected beam angles were within ±15° and resulting plans showed minimal dosimetric differences. In cases 7-9 AI beam angles reduced max dose while keeping organ-at-risk dose within ±2 Gy. In case 10 the AI chosen beam angles reduced cord Dmax by 9.1 Gy but increased left eye Dmax by 4.9 Gy, other organs-at-risk showed minimal differences. Gantry beam angle selection was automated by a newly developed AI model and tested on 10 brain IMPT patients. The comparison showed that the AI often chooses similar beam angles to the human planners. When there are differences, dosimetric analysis demonstrated that plans created from AI beam angles have at least the same quality as the human ones. Results motivate further research into this approach showing the AI being a promising tool to fill a current gap in the strive for automating proton treatment planning, increasing planning efficiency and potentially quality.

Multivariate Emulation of Kilometer-Scale Numerical Weather Predictions with Generative Adversarial Networks: A Proof of Concept

Abstract Emulating numerical weather prediction (NWP) model outputs is important to compute large datasets of weather fields in an efficient way. The purpose of the present paper is to investigate the ability of generative adversarial networks (GANs) to emulate distributions of multivariate outputs (10-m wind and 2-m temperature) of a kilometer-scale NWP model. For that purpose, a residual GAN architecture, regularized with spectral normalization, is trained against a kilometer-scale dataset from the AROME Ensemble Prediction System (AROME-EPS). A wide range of metrics is used for quality assessment, including pixelwise and multiscale Earth-mover distances, spectral analysis, and correlation length scales. The use of wavelet-based scattering coefficients as meaningful metrics is also presented. The GAN generates samples with good distribution recovery and good skill in average spectrum reconstruction. Important local weather patterns are reproduced with a high level of detail, while the joint generation of multivariate samples matches the underlying AROME-EPS distribution. The different metrics introduced describe the GAN’s behavior in a complementary manner, highlighting the need to go beyond spectral analysis in generation quality assessment. An ablation study then shows that removing variables from the generation process is globally beneficial, pointing at the GAN limitations to leverage cross-variable correlations. The role of absolute positional bias in the training process is also characterized, explaining both accelerated learning and quality-diversity trade-off in the multivariate emulation. These results open perspectives about the use of GAN to enrich NWP ensemble approaches, provided that the aforementioned positional bias is properly controlled.

Graphic Design of 3D Animation Scenes Based on Deep Learning and Information Security Technology

This paper aims to use the improved Generative Adversarial Network (GAN) model for Three Dimensional (3D) animation graphic design, improve the efficiency of 3D animation graphic design, and promote the accuracy of model recognition. It acquires 3D animated scene color images from different perspectives. This paper performs 3D visualization through point clouds, outputs high-quality point cloud results, and uses Convolutional Neural Network (CNN), Earth-Mover (EM) distance, and Least Squares Method (LSM) to improve the GAN model. Finally, the effectiveness of the improved GAN in the graphic design of 3D animation scenes and the effects of different improved models in generating 3D animation scene images are analyzed. The results show that the computational loss amplitude of the improved GAN model using Label Smoothing processing deep convolutional neural network is between [2, 3]. The generator loss variation is smaller, and the image quality of the generated 3D animation scene is gradually improved. The training process of the LSM-improved model is more stable, and the loss value is lower than that of the EM distance improved model. The loss value of the generator is [0.3,0.5], and the loss value of the discriminator is [0.1,0.2]. The Inception score of the LSM-improved model is 0.0297 higher than that of the CNN-improved model and the EM distance improved model and 0.0198 higher than that of the GAN model.

Feature-level consistency regularized Semi-supervised scheme with data augmentation for intelligent fault diagnosis under small samples

Intelligent fault diagnosis based on machine learning has yielded a wealth of research results. However, fault diagnosis under small samples is still a challenging problem due to the difficulty of collecting fault data of machines in engineering scenarios. To address this problem, this paper proposes a feature-level consistency regularized semi-supervised scheme with data augmentation for fault diagnosis of machines. In the proposed method, generative adversarial networks generate unlabeled data to augment the limited training set. The fault identification network in the proposed method is trained by the augmented data and a small amount of real data collected from machines in a semi-supervised learning manner. We construct a distance evaluation metric based on the Earth Mover's Distance (EMD) and further design a novel feature consistency regularization module, which helps the fault identification network learn robust fault features by minimizing the EMD between generated data features and real data features. To verify the effectiveness of the proposed model, two mechanical fault simulation experiments were carried out. The experimental results show that the proposed method achieves high fault diagnosis accuracy using only a small amount of training data samples.