Large-scale Training Samples Research Articles

Abstract 4800: Reducing health disparities for prostate adenocarcinoma by integrating multi-omics data via a multi-modal transfer learning approach

Abstract Prostate adenocarcinoma (PRAD) is the most common subtype of prostate cancer, which is the second leading cause of cancer death in American men. Although significant progress has been made to improve PRAD prognosis, existing studies indicated that Black American men have disproportionately high incidence and mortality rate in prostate cancer compared with non-Hispanic White American men. With artificial intelligence (AI) and machine learning (ML) being increasingly applied to PRAD research and clinical decision-making, PRAD data disparities would introduce bias to AI/ML models and further enhance negative impacts on healthcare towards underrepresented groups. Transfer learning has shown potential to reduce racial disparities in PRAD. However, its performance may be deteriorated due to: (1) its model requires large-scale training samples which are difficult to obtain in clinical settings, and (2) it only uses single-omics data without integrating multi-omics information. To address these concerns, we propose to develop a multi-modal transfer learning model to integrate multi-omics data for reducing health disparities. Specifically, we first investigated two multi-modal ensemble methods, Pearson Correlation Coefficient (PCC) based patient-pairwise similarity, and variational autoencoder (VAE) to integrate different types of omics data. Then, we leveraged a transfer learning model based on domain adaptation to pre-train the model on the majority group (White Americans) and fine-tune the model using the minority group (Black Americans). To further address the imbalanced data among ethnic groups, we explored implementing a data augmentation method, Synthetic Minority Oversampling Technique (SMOTE), to increase the size of minority group data. We evaluated our model on multi-omics data (mRNA, miRNA and methylation) of PRAD from The Cancer Genome Atlas (TCGA) database. Results suggested that our proposed approach achieved better performance of reducing health disparities for Black Americans compared with the mixture model and the independent model as well as the conventional transfer learning model in predicting progression-free interval (PFI) prognosis for PRAD patients. Furthermore, we also demonstrated that SMOTE alleviated the data imbalanced problem and improved the performance of the prognosis classification for ethnic minority groups. In summary, our results demonstrated that our proposed multi-modal transfer learning approach could effectively reduce the performance gap between majority and minority groups in PRAD prognosis, thereby mitigating health disparities. We expect that our proposed multi-model transfer learning framework can be customized and extensible to reduce health disparities in other types of cancer. Citation Format: Lusheng Li, Jieqiong Wang, Shibiao Wan. Reducing health disparities for prostate adenocarcinoma by integrating multi-omics data via a multi-modal transfer learning approach [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 4800.

Self-Supervised Medical Image Denoising Based on WISTA-Net for Human Healthcare in Metaverse.

Medical image processing plays an important role in the interaction of real world and metaverse for healthcare. Self-supervised denoising based on sparse coding methods, without any prerequisite on large-scale training samples, has been attracting extensive attention for medical image processing. Whereas, existing self-supervised methods suffer from poor performance and low efficiency. In this paper, to achieve state-of-the-art denoising performance on the one hand, we present a self-supervised sparse coding method, named the weighted iterative shrinkage thresholding algorithm (WISTA). It does not rely on noisy-clean ground-truth image pairs to learn from only a single noisy image. On the other hand, to further improve denoising efficiency, we unfold the WISTA to construct a deep neural network (DNN) structured WISTA, named WISTA-Net. Specifically, in WISTA, motivated by the merit of the lp-norm, WISTA-Net has better denoising performance than the classical orthogonal matching pursuit (OMP) algorithm and the ISTA. Moreover, leveraging the high-efficiency of DNN structure in parameter updating, WISTA-Net outperforms the compared methods in denoising efficiency. In detail, for a 256 by 256 noisy image, the running time of WISTA-Net is 4.72 s on the CPU, which is much faster than WISTA, OMP, and ISTA by 32.88 s, 13.06 s, and 6.17 s, respectively.

Supervised spectral feature learning for fine-grained classification in small data set

Fine-grained image classification is a challenging task due to the small inter-class variance, the large intra-class difference, and the small training data. Traditional methods typically rely on large-scale training samples with annotated part annotations, making them costly and severely limiting their application area. In this paper, we propose an effective and weakly supervised fine-grained classification framework. In this framework, a discriminative class-specific spectral feature is learned by intra-class spectral coupling and inter-class spectral decoupling under the weak supervision of image-level category labels, and then the new input images are classified based on the learned class-specific spectral feature. Different from existing strong supervised methods, the proposed technique creatively combines weak supervision of the image-level category labels with unsupervised spectral graph decomposition, not relying on large-scale training samples with dense part annotations, which are heavily labor-consuming. The performance of the proposed methods has been verified on four kinds of typical datasets: the JAFFE dataset, the Yale database, the UCI-CMU face database, and the neural foramina dataset. The satisfactory classification results have been achieved by the proposed method in expression recognition on the JAFFE dataset (with a mean accuracy of 95.31%), face recognition on the Yale database (with a mean accuracy of 98.79%), object recognition on the UCI-CMU face database (with a mean accuracy of 96.96%), and disease grading on the neural foramina dataset (with a mean accuracy of 92.09%). Compared with most state-of-the-art methods, the proposed method has superior classification performance in the small data set.

Laplace-Domain Hybrid Distribution Model Based FDIA Attack Sample Generation in Smart Grids

False data injection attack (FDIA) is a deliberate modification of measurement data collected by the power grid using vulnerabilities in power grid state estimation, resulting in erroneous judgments made by the power grid control center. As a symmetrical defense scheme, FDIA detection usually uses machine learning methods to detect attack samples. However, existing detection models for FDIA typically require large-scale training samples, which are difficult to obtain in practical scenarios, making it difficult for detection models to achieve effective detection performance. In light of this, this paper proposes a novel FDIA sample generation method to construct large-scale attack samples by introducing a hybrid Laplacian model capable of accurately fitting the distribution of data changes. First, we analyze the large-scale power system sensing measurement data and establish the data distribution model of symmetric Laplace distribution. Furthermore, a hybrid Laplace-domain symmetric distribution model with multi-dimensional component parameters is constructed, which can induce a deliberate deviation in the state estimation from its safe value by injecting into the power system measurement. Due to the influence of the multivariate parameters of the hybrid Laplace-domain distribution model, the sample deviation generated by this model can not only obtain an efficient attack effect, but also effectively avoid the recognition of the FDIA detection model. Extensive experiments are carried out over IEEE 14-bus and IEEE 118-bus test systems. The corresponding results unequivocally demonstrate that our proposed attack method can quickly construct large-scale FDIA attack samples and exhibit significantly higher resistance to detection by state-of-the-art detection models, while also offering superior concealment capabilities compared to traditional FDIA approaches.

A large scale training sample database system for intelligent interpretation of remote sensing imagery

ABSTRACT Artificial Intelligence (AI) Machine Learning (ML) technologies, particularly Deep Learning (DL), have demonstrated significant potential in the interpretation of Remote Sensing (RS) imagery, covering tasks such as scene classification, object detection, land-cover/land-use classification, change detection, and multi-view stereo reconstruction. Large-scale training samples are essential for ML/DL models to achieve optimal performance. However, the current organization of training samples is ad-hoc and vendor-specific, lacking an integrated approach that can effectively manage training samples from different vendors to meet the demands of various RS AI tasks. This article proposes a solution to address these challenges by designing and implementing LuoJiaSET, a large-scale training sample database system for intelligent interpretation of RS imagery. LuoJiaSET accommodates over five million training samples, providing support for cross-dataset queries and serving as a comprehensive training data store for RS AI model training and calibration. It overcomes challenges related to label semantic categories, structural heterogeneity in label representation, and interoperable data access.

Bayesian self-supervised learning allying with Transformer powered compressed sensing imaging

Deep learning based compressed sensing imaging algorithms usually adopt supervised learning mechanism, which generally require massive high-quality images paired with the measured data in the training process. However, clear images are difficult to obtain in some specific application scenarios. In view of this, we propose a self-attention powered multi-scale compressed sensing imaging network named BNN-SANet with the Bayesian self-supervised learning manner in this paper. BNN-SANet can automatically learn effective feature representation from the measured data and reconstruct high-quality images without large-scale training samples. In particular, the self-attention based Transformer employed in BNN-SANet is effective to exploit global and long-range dependencies. Multi-scale dual path design in BNN-SANet performs well in extracting the semantic information at different scales to further improve the CS reconstruction quality. Experimental results illustrate that the proposed approach is substantially superior to previous compressed sensing imaging methods in terms of Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) and visual metrics.

Few-Shot Remote Sensing Image Scene Classification Based on Metric Learning and Local Descriptors

Scene classification is a critical technology to solve the challenges of image search and image recognition. It has become an indispensable and challenging research topic in the field of remote sensing. At present, most scene classifications are solved by deep neural networks. However, existing methods require large-scale training samples and are not suitable for actual scenarios with only a few samples. For this reason, a framework based on metric learning and local descriptors (MLLD) is proposed to enhance the classification effect of remote sensing scenes on the basis of few-shot. Specifically, MLLD adopts task-level training that is carried out through meta-learning, and meta-knowledge is learned to improve the model’s ability to recognize different categories. Moreover, Manifold Mixup is introduced by MLLD as a feature processor for the hidden layer of deep neural networks to increase the low confidence space for smoother decision boundaries and simpler hidden layer representations. In the end, a learnable metric is introduced; the nearest category of the image is matched by measuring the similarity of local descriptors. Experiments are conducted on three public datasets: UC Merced, WHU-RS19, and NWPU-RESISC45. Experimental results show that the proposed scene classification method can achieve the most advanced results on limited datasets.

Zero-Shot Blind Learning for Single-Image Super-Resolution

Deep convolutional neural networks (DCNNs) have manifested significant performance gains for single-image super-resolution (SISR) in the past few years. Most of the existing methods are generally implemented in a fully supervised way using large-scale training samples and only learn the SR models restricted to specific data. Thus, the adaptation of these models to real low-resolution (LR) images captured under uncontrolled imaging conditions usually leads to poor SR results. This study proposes a zero-shot blind SR framework via leveraging the power of deep learning, but without the requirement of the prior training using predefined imaged samples. It is well known that there are two unknown data: the underlying target high-resolution (HR) images and the degradation operations in the imaging procedure hidden in the observed LR images. Taking these in mind, we specifically employed two deep networks for respectively modeling the priors of both the target HR image and its corresponding degradation kernel and designed a degradation block to realize the observation procedure of the LR image. Via formulating the loss function as the approximation error of the observed LR image, we established a completely blind end-to-end zero-shot learning framework for simultaneously predicting the target HR image and the degradation kernel without any external data. In particular, we adopted a multi-scale encoder–decoder subnet to serve as the image prior learning network, a simple fully connected subnet to serve as the kernel prior learning network, and a specific depthwise convolutional block to implement the degradation procedure. We conducted extensive experiments on several benchmark datasets and manifested the great superiority and high generalization of our method over both SOTA supervised and unsupervised SR methods.

Feature Consistency Constraints-Based CNN for Landsat Land Cover Mapping

The cascade of convolution layers and the end-to-end training process facilitate CNN feature extraction and transmission, and promote the success of CNN in image processing. However, the drawback of heavily relying on large-scale high-quality training samples restricts its applications. To avoid costly and unrealistic manual annotations for large-scale remote sensing images, existing land cover maps are considered as an alternative to manual annotations, in which noisy labels are inevitable. To alleviate the impact of noisy labels, this paper proposes to improve the consistency feature learning ability of CNNs as a feasible solution in practical land cover mapping. Firstly, an intraclass feature consistency constraint is introduced to maintain the consistency of CNN feature maps for the same class. Then, an inter-iteration feature consistency constraint is employed to guide the network to learn features that are consistent with the whole underlying distribution inside a mini-batch. These two feature consistency constraints work in a cooperative and complementary manner with the traditional cross-entropy, and together improve the consistency feature learning ability of the proposed Feature Consistency Network (FCNet). Experimental results demonstrate the effectiveness of the proposed FCNet. Extensive experiments on different network structures validate the generalization of the proposed feature consistency constraints.

Parameters tuning of multi-model database based on deep reinforcement learning

As we all know, the performance of database management system is directly linked to a vast array of knobs, which control various aspects of system operation, ranging from memory and thread counts settings to I/O optimization. Improper settings of configuration parameters are shown to have detrimental effects on performance, reliability and availability of the overall database management system. This is also true for multi-model databases, which use a single platform to support multiple data models. Existing approaches for automatic DBMS knobs tuning are not directly applicable to multi-model databases due to the diversity of multi-model database instances and workloads. Firstly, in cloud environment, they have difficulty adapting to changing environments and diverse workloads. Secondly, they rely on large-scale high-quality training samples that are difficult to obtain. Finally, they focus primarily on throughput metrics, ignoring tuning requirements for resource utilization. Therefore, in this paper, we propose a multi-model database configuration parameters tuning solution named MMDTune. It selects influential parameters, recommends the optimal configurations in a high-dimensional continuous space. For different workloads, the TD3 algorithm is improved to generate reasonable parameter adjustment plans according to the internal state of the multi-model databases. We conduct extensive experiments under 5 different workloads on real cloud databases to evaluate MMDTune. Experimental results show that MMDTune adapts well to a new hardware environment or workloads, and significantly outperforms the representative tuning tools, such as OtterTune, CDBTune.

Deep Metric Learning-Based for Multi-Target Few-Shot Pavement Distress Classification

Pavement distress detection is of great significance for road maintenance and to ensure road safety. At present, detection methods based on deep learning have achieved outstanding performance in related fields. However, these methods require large-scale training samples. For pavement distress detection, it is difficult to collect more images with pavement distress, and the types of pavement diseases are increasing with time, so it is impossible to ensure sufficient pavement distress samples to train the supervised deep model. In this article, we propose a new few-shot pavement distress detection method based on metric learning, which can effectively learn new categories from a few labeled samples. In this article, we adopt the backend network (ResNet18) to extract multilevel feature information from the base classes and then send the extracted features into the metric module. In the metric module, we introduce the attention mechanism to learn the feature attributes of “what” and “where” and focus the model on the desired characteristics. We also introduce a new metric loss function to maximize the distance between different categories while minimizing the distance between the same categories. In the testing stage, we calculate the cosine similarity between the support set and query set to complete novel category detection. The experimental results show that the proposed method significantly outperforms several benchmarking methods on the pavement distress dataset (the classification accuracies of 5-way 1-shot and 5-way 5-shot are 77.20% and 87.28%, respectively).

Weakly Supervised Semantic and Attentive Data Mixing Augmentation for Fine-Grained Visual Categorization

As a key factor, the availability of large-scale training samples determines the improvement of visual performance. However, the size of Fine-Grained Visual Categorization (FGVC) datasets is always limited. Therefore, overfitting as an issue in FGVC-related training needs to be solved. Data mixing augmentation is a widely-used data augmentation method. In most of the recently proposed data mixing augmentation methods, random patch selection may generate meaningless training samples and result in model instability during the training process. This paper proposes a data mixing augmentation strategy termed Semantic and Attentive Data Mixing (SADMix) to select semantic patches for the generation of new training samples. In SADMix, a certain number of critical regions are localized according to convolutional activations. An image patch is selected from these localized regions for the generation of new training samples. The size, aspect ratio, and center location of these image patches are changed according to the random values from a beta distribution. These image patches with semantic information are used to mix two training images. According to the class activation map (CAM), training images and their labels are mixed proportionally to generate new mixed training samples and the corresponding labels. The proposed SADMix is tested on three fine-grained datasets, which are CUB-200-2011, FGVC Aircraft, and Stanford Cars, respectively. The experimental results confirm the effectiveness of the proposed SADMix.

Distinguishing between natural and recolored images via lateral chromatic aberration

With the development of image colorization technique, the recolored images (RIs) become more and more authentic, making it very difficult to visually distinguish from natural images (NIs). Recently, researchers have proposed the detection methods towards recolored images. However, the current detection still has limitations such as poor generalization, large-scale training samples, high-dimensional features for training, and high computation cost. To address those issues, this paper proposes a novel method based on the lateral chromatic aberration (LCA) inconsistency and its statistical differences. Generally, RIs have fewer numbers of LCA characteristics than that of NIs, that inspire us to design the classifier for distinguishing two types of images. In particular, we propose to adopt very low 5-dimensional features to feed a classical SVM mechanism. The baseline ImageNet and Oxford datasets are used to verify the effectiveness of the proposed method, in which the performance of our proposed method rivals the prior arts.

Hbox {CDBTune}^{+}: An efficient deep reinforcement learning-based automatic cloud database tuning system

Configuration tuning is vital to optimize the performance of a database management system (DBMS). It becomes more tedious and urgent for cloud databases (CDB) due to diverse database instances and query workloads, which make the job of a database administrator (DBA) very difficult. Existing solutions for automatic DBMS configuration tuning have several limitations. Firstly, they adopt a pipelined learning model but cannot optimize the overall performance in an end-to-end manner. Secondly, they rely on large-scale high-quality training samples which are hard to obtain. Thirdly, existing approaches cannot recommend reasonable configurations for a large number of knobs to tune whose potential values live in such high-dimensional continuous space. Lastly, in cloud environments, existing approaches can hardly cope with the changes of hardware configurations and workloads, and have poor adaptability. To address these challenges, we design an end-to-end automatic CDB tuning system, {texttt {CDBTune}}^{+}, using deep reinforcement learning (RL). {texttt {CDBTune}}^{+} utilizes the deep deterministic policy gradient method to find the optimal configurations in a high-dimensional continuous space. {texttt {CDBTune}}^{+} adopts a trial-and-error strategy to learn knob settings with a limited number of samples to accomplish the initial training, which alleviates the necessity of collecting a massive amount of high-quality samples. {texttt {CDBTune}}^{+} adopts the reward-feedback mechanism in RL instead of traditional regression, which enables end-to-end learning and accelerates the convergence speed of our model and improves the efficiency of online tuning. Besides, we propose effective techniques to improve the training and tuning efficiency of {texttt {CDBTune}}^{+} for practical usage in a cloud environment. We conducted extensive experiments under 7 different workloads on real cloud databases to evaluate {texttt {CDBTune}}^{+}. Experimental results showed that {texttt {CDBTune}}^{+} adapts well to a new hardware environment or workload, and significantly outperformed the state-of-the-art tuning tools and DBA experts.

MetaStore: A Task-adaptative Meta-learning Model for Optimal Store Placement with Multi-city Knowledge Transfer

Optimal store placement aims to identify the optimal location for a new brick-and-mortar store that can maximize its sale by analyzing and mining users’ preferences from large-scale urban data. In recent years, the expansion of chain enterprises in new cities brings some challenges because of two aspects: (1) data scarcity in new cities, so most existing models tend to not work (i.e., overfitting), because the superior performance of these works is conditioned on large-scale training samples; (2) data distribution discrepancy among different cities, so knowledge learned from other cities cannot be utilized directly in new cities. In this article, we propose a task-adaptative model-agnostic meta-learning framework, namely, MetaStore, to tackle these two challenges and improve the prediction performance in new cities with insufficient data for optimal store placement, by transferring prior knowledge learned from multiple data-rich cities. Specifically, we develop a task-adaptative meta-learning algorithm to learn city-specific prior initializations from multiple cities, which is capable of handling the multimodal data distribution and accelerating the adaptation in new cities compared to other methods. In addition, we design an effective learning strategy for MetaStore to promote faster convergence and optimization by sampling high-quality data for each training batch in view of noisy data in practical applications. The extensive experimental results demonstrate that our proposed method leads to state-of-the-art performance compared with various baselines.

Discussion of Training Samples-Oreinted Sampling Methods for Data-driven Power Systems Analysis

Data-driven methods such as deep neural networks (DNNs) and Gaussian processes (GPs), have been a promising way to achieve the balance of the calculation accuracy and efficiency in power system analysis. For instance, studies are utilizing the DNN-based method to achieve the fast and accurate calculation of probabilistic power flow. These methods rely on the sample data for training, which is mainly obtained by sampling. However, there is still no guidance on how to select the suitable sampling method to effectively generate the representative training samples. To address this problem, this paper explores the impact of different sampling methods on data-driven methods through theoretical analysis and simulation. Then some reasonable conclusions are drawn in this paper: i) Random sampling (RS) is suitable for data-driven methods that require large-scale training samples. ii) Latin hypercube sampling (LHS) is suitable for data-driven methods that only work well under small-scale training samples. iii) The assumption that the training samples generated by uniform sampling (US), which covers all sampling intervals, could not lead to good performance. Finally, simulations are implemented on the IEEE 39-bus system and the results demonstrate the effectiveness of the proposed conclusions

A cost-effective manufacturing process recognition approach based on deep transfer learning for CPS enabled shop-floor

The rapid development of the industrial Internet of Things has promoted manufacturing to develop towards the cyber-physical system, of which highly accurate process recognition plays an important role in achieving proactive monitoring of intelligent manufacturing process. Compared to the traditional handcrafted feature-based method, deep model owns convenience in terms of extracting feature automatically for the recognition. However, training a deep model is time-consuming and also requires large-scale training samples. To solve these problems and obtain high accuracy in the meanwhile, a deep transfer learning-based manufacturing process recognition approach is proposed in this study. A pre-trained model based on a convolutional neural network is used to extract low dimensional features followed by a fine-tuning process to target the specific process recognition task. Experimental verification of two datasets was conducted to demonstrate this cost-effective method. The results showed the proposed method can get better accuracy with less training time and fewer training samples.

Image Tampering Localization Using a Dense Fully Convolutional Network

The emergence of powerful image editing software has substantially facilitated digital image tampering, leading to many security issues. Hence, it is urgent to identify tampered images and localize tampered regions. Although much attention has been devoted to image tampering localization in recent years, it is still challenging to perform tampering localization in practical forensic applications. The reasons include the difficulty of learning discriminative representations of tampering traces and the lack of realistic tampered images for training. Since Photoshop is widely used for image tampering in practice, this paper attempts to address the issue of tampering localization by focusing on the detection of commonly used editing tools and operations in Photoshop. In order to well capture tampering traces, a fully convolutional encoder-decoder architecture is designed, where dense connections and dilated convolutions are adopted for achieving better localization performance. In order to effectively train a model in the case of insufficient tampered images, we design a training data generation strategy by resorting to Photoshop scripting, which can imitate human manipulations and generate large-scale training samples. Extensive experimental results show that the proposed approach outperforms state-of-the-art competitors when the model is trained with only generated images or fine-tuned with a small amount of realistic tampered images. The proposed method also has good robustness against some common post-processing operations.

Knowledge-driven Egocentric Multimodal Activity Recognition

Recognizing activities from egocentric multimodal data collected by wearable cameras and sensors, is gaining interest, as multimodal methods always benefit from the complementarity of different modalities. However, since high-dimensional videos contain rich high-level semantic information while low-dimensional sensor signals describe simple motion patterns of the wearer, the large modality gap between the videos and the sensor signals raises a challenge for fusing the raw data. Moreover, the lack of large-scale egocentric multimodal datasets due to the cost of data collection and annotation processes makes another challenge for employing complex deep learning models. To jointly deal with the above two challenges, we propose a knowledge-driven multimodal activity recognition framework that exploits external knowledge to fuse multimodal data and reduce the dependence on large-scale training samples. Specifically, we design a dual-GCLSTM (Graph Convolutional LSTM) and a multi-layer GCN (Graph Convolutional Network) to collectively model the relations among activities and intermediate objects. The dual-GCLSTM is designed to fuse temporal multimodal features with top-down relation-aware guidance. In addition, we apply a co-attention mechanism to adaptively attend to the features of different modalities at different timesteps. The multi-layer GCN aims to learn relation-aware classifiers of activity categories. Experimental results on three publicly available egocentric multimodal datasets show the effectiveness of the proposed model.

Topology-guided cyclic brain connectivity generation using geometric deep learning

BackgroundThere is a growing need for analyzing medical data such as brain connectomes. However, the unavailability of large-scale training samples increases risks of model over-fitting. Recently, deep learning (DL) architectures quickly gained momentum in synthesizing medical data. However, such frameworks are primarily designed for Euclidean data (e.g., images), overlooking geometric data (e.g., brain connectomes). A few existing geometric DL works that aimed to predict a target brain connectome from a source one primarily focused on domain alignment and were agnostic to preserving the connectome topology. New methodTo address the above limitations, firstly, we adapt the graph translation generative adversarial network (GT GAN) architecture to brain connectomic data. Secondly, we extend the baseline GT GAN to a cyclic graph translation (CGT) GAN, allowing bidirectional brain network translation between the source and target views. Finally, to preserve the topological strength of brain regions of interest (ROIs), we impose a topological strength constraint on the CGT GAN learning, thereby introducing CGTS GAN architecture. Comparison with existing methodsWe compared CGTS with graph translation methods and its ablated versions. ResultsOur deep graph network outperformed the baseline comparison method and its ablated versions in mean squared error (MSE) using multiview autism spectrum disorder connectomic dataset. ConclusionWe designed a topology-aware bidirectional brain connectome synthesis framework rooted in geometric deep learning, which can be used for data augmentation in clinical diagnosis.