Multimodal Deep Neural Network Research Articles

SG-Fusion: A swin-transformer and graph convolution-based multi-modal deep neural network for glioma prognosis

The integration of morphological attributes extracted from histopathological images and genomic data holds significant importance in advancing tumor diagnosis, prognosis, and grading. Histopathological images are acquired through microscopic examination of tissue slices, providing valuable insights into cellular structures and pathological features. On the other hand, genomic data provides information about tumor gene expression and functionality. The fusion of these two distinct data types is crucial for gaining a more comprehensive understanding of tumor characteristics and progression. In the past, many studies relied on single-modal approaches for tumor diagnosis. However, these approaches had limitations as they were unable to fully harness the information from multiple data sources. To address these limitations, researchers have turned to multi-modal methods that concurrently leverage both histopathological images and genomic data. These methods better capture the multifaceted nature of tumors and enhance diagnostic accuracy. Nonetheless, existing multi-modal methods have, to some extent, oversimplified the extraction processes for both modalities and the fusion process. In this study, we presented a dual-branch neural network, namely SG-Fusion. Specifically, for the histopathological modality, we utilize the Swin-Transformer structure to capture both local and global features and incorporate contrastive learning to encourage the model to discern commonalities and differences in the representation space. For the genomic modality, we developed a graph convolutional network based on gene functional and expression level similarities. Additionally, our model integrates a cross-attention module to enhance information interaction and employs divergence-based regularization to enhance the model’s generalization performance. Validation conducted on glioma datasets from the Cancer Genome Atlas unequivocally demonstrates that our SG-Fusion model outperforms both single-modal methods and existing multi-modal approaches in both survival analysis and tumor grading.

Identifying chaotic dynamics in noisy time series through multimodal deep neural networks

Abstract Chaos detection is the problem of identifying whether a series of measurements is being sampled from an underlying set of chaotic dynamics. The unavoidable presence of measurement noise significantly affects the performance of chaos detectors, as discerning chaotic dynamics from stochastic signals becomes more challenging. This paper presents a computationally efficient multi-modal deep neural network tailored for chaos detection by combining information coming from the analysis of time series, recurrence plots and spectrograms. The proposed approach is the first one suitable for multi-class classification of chaotic systems while being robust with respect to measurement noise, and is validated on a dataset of 15 different chaotic and non-chaotic dynamics subject to white, pink or brown coloured noise.

AQINet: A multimodal deep convolutional neural network to predict Air Quality Index via satellite imagery and meteorological data

AQINet: A multimodal deep convolutional neural network to predict Air Quality Index via satellite imagery and meteorological data

A novel deep learning framework for accurate melanoma diagnosis integrating imaging and genomic data for improved patient outcomes.

Melanoma is one of the most malignant forms of skin cancer, with a high mortality rate in the advanced stages. Therefore, early and accurate detection of melanoma plays an important role in improving patients' prognosis. Biopsy is the traditional method for melanoma diagnosis, but this method lacks reliability. Therefore, it is important to apply new methods to diagnose melanoma effectively. This study presents a new approach to classify melanoma using deep neural networks (DNNs) with combined multiple modal imaging and genomic data, which could potentially provide more reliable diagnosis than current medical methods for melanoma. We built a dataset of dermoscopic images, histopathological slides and genomic profiles. We developed a custom framework composed of two widely established types of neural networks for analysing image data Convolutional Neural Networks (CNNs) and networks that can learn graph structure for analysing genomic data-Graph Neural Networks. We trained and evaluated the proposed framework on this dataset. The developed multi-modal DNN achieved higher accuracy than traditional medical approaches. The mean accuracy of the proposed model was 92.5% with an area under the receiver operating characteristic curve of 0.96, suggesting that the multi-modal DNN approach can detect critical morphologic and molecular features of melanoma beyond the limitations of traditional AI and traditional machine learning approaches. The combination of cutting-edge AI may allow access to a broader range of diagnostic data, which can allow dermatologists to make more accurate decisions and refine treatment strategies. However, the application of the framework will have to be validated at a larger scale and more clinical trials need to be conducted to establish whether this novel diagnostic approach will be more effective and feasible.

TinyM 2 Net-V2: A Compact Low-power Software Hardware Architecture for M ulti m odal Deep Neural Networks

With the evaluation of Artificial Intelligence (AI), there has been a resurgence of interest in how to use AI algorithms on low-power embedded systems to broaden potential use cases of the Internet of Things (IoT). To mimic multimodal human perception, multimodal deep neural networks (M-DNN) have recently become very popular with the classification task due to their impressive performance for computer vision and audio processing tasks. This article presents TinyM 2 Net-V2 —a compact low-power software hardware architecture for m ulti m odal deep neural networks for resource-constrained tiny devices. To compress the models to implement on tiny devices, cyclicly sparsification and hybrid quantization (4-bits weights and 8-bits activations) methods are used. Although model compression techniques are an active research area, we are the first to demonstrate their efficacy for multimodal deep neural networks, using cyclicly sparsification and hybrid quantization of weights/activations. TinyM 2 Net-V2 shows that even a tiny multimodal deep neural network model can improve the classification accuracy more than that of any unimodal counterparts. Parameterized M-DNN model architecture was designed to be evaluated in two different case-studies: vehicle detection from multimodal images and audios and COVID-19 detection from multimodal audio recordings. The most compressed TinyM 2 Net-V2 achieves 92.5% COVID-19 detection accuracy (6.8% improvement from the unimodal full precision model) and 90.6% vehicle classification accuracy (7.7% improvement from the unimodal full precision model). A parameterized and flexible FPGA hardware accelerator was designed as well for TinyM 2 Net-V2 models. To the best of our knowledge, this is the first work accelerating multimodal deep neural network models on low-power Artix-7 FPGA hardware. We achieved energy efficiency of 9.04 GOP/s/W and 15.38 GOP/s/W for case-study 1 and case-study 2, respectively, which is comparable to the state-of-the-art results. Finally, we compared our tiny FPGA hardware implementation results with off-the-shelf resource-constrained devices and showed our implementation is faster and consumed less power compared to the off-the-shelf resource-constrained devices.

Multimodal protein representation learning and target-aware variational auto-encoders for protein-binding ligand generation

Without knowledge of specific pockets, generating ligands based on the global structure of a protein target plays a crucial role in drug discovery as it helps reduce the search space for potential drug-like candidates in the pipeline. However, contemporary methods require optimizing tailored networks for each protein, which is arduous and costly. To address this issue, we introduce TargetVAE, a target-aware variational auto-encoder that generates ligands with desirable properties including high binding affinity and high synthesizability to arbitrary target proteins, guided by a multimodal deep neural network built based on geometric and sequence models, named Protein Multimodal Network (PMN), as the prior for the generative model. PMN unifies different representations of proteins (e.g. primary structure—sequence of amino acids, 3D tertiary structure, and residue-level graph) into a single representation. Our multimodal architecture learns from the entire protein structure and is able to capture their sequential, topological, and geometrical information by utilizing language modeling, graph neural networks, and geometric deep learning. We showcase the superiority of our approach by conducting extensive experiments and evaluations, including predicting protein-ligand binding affinity in the PBDBind v2020 dataset as well as the assessment of generative model quality, ligand generation for unseen targets, and docking score computation. Empirical results demonstrate the promising and competitive performance of our proposed approach. Our software package is publicly available at https://github.com/HySonLab/Ligand_Generation.

Exploratory drug discovery in breast cancer patients: A multimodal deep learning approach to identify novel drug candidates targeting RTK signaling

Breast cancer, a highly formidable and diverse malignancy predominantly affecting women globally, poses a significant threat due to its intricate genetic variability, rendering it challenging to diagnose accurately. Various therapies such as immunotherapy, radiotherapy, and diverse chemotherapy approaches like drug repurposing and combination therapy are widely used depending on cancer subtype and metastasis severity. Our study revolves around an innovative drug discovery strategy targeting potential drug candidates specific to RTK signalling, a prominently targeted receptor class in cancer. To accomplish this, we have developed a multimodal deep neural network (MM-DNN) based QSAR model integrating omics datasets to elucidate genomic, proteomic expression data, and drug responses, validated rigorously. The results showcase an R2 value of 0.917 and an RMSE value of 0.312, affirming the model's commendable predictive capabilities. Structural analogs of drug molecules specific to RTK signalling were sourced from the PubChem database, followed by meticulous screening to eliminate dissimilar compounds. Leveraging the MM-DNN-based QSAR model, we predicted the biological activity of these molecules, subsequently clustering them into three distinct groups. Feature importance analysis was performed. Consequently, we successfully identified prime drug candidates tailored for each potential downstream regulatory protein within the RTK signalling pathway. This method makes the early stages of drug development faster by removing inactive compounds, providing a hopeful path in combating breast cancer.

Different gait combinations based on multi-modal deep CNN architectures

Gait recognition is the process of identifying a person from a distance based on their walking patterns. However, the recognition rate drops significantly under cross-view angle and appearance-based variations. In this study, the effectiveness of the most well-known gait representations in solving this problem is investigated based on deep learning. For this purpose, a comprehensive performance evaluation is performed by combining different modalities, including silhouettes, optical flows, and concatenated image of the Gait Energy Image (GEI) head and leg region, with GEI itself. This evaluation is carried out across different multimodal deep convolutional neural network (CNN) architectures, namely fine-tuned EfficientNet-B0, MobileNet-V1, and ConvNeXt-base models. These models are trained separately on GEIs, silhouettes, optical flows, and concatenated image of GEI head and leg regions, and then extracted GEI features are fused in pairs with other extracted modality features to find the most effective gait combination. Experimental results on the two different datasets CASIA-B and Outdoor-Gait show that the concatenated image of GEI head and leg regions significantly increased the recognition rate of the networks compared to other modalities. Moreover, this modality demonstrates greater robustness under varied carrying (BG) and clothing (CL) conditions compared to optical flows (OF) and silhouettes (SF). Codes available at https://github.com/busrakckugurlu/Different-gait-combinations-based-on-multi-modal-deep-CNN-architectures.git

Beyond images: an integrative multi-modal approach to chest x-ray report generation.

Image-to-text radiology report generation aims to automatically produce radiology reports that describe the findings in medical images. Most existing methods focus solely on the image data, disregarding the other patient information accessible to radiologists. In this paper, we present a novel multi-modal deep neural network framework for generating chest x-rays reports by integrating structured patient data, such as vital signs and symptoms, alongside unstructured clinical notes. We introduce a conditioned cross-multi-head attention module to fuse these heterogeneous data modalities, bridging the semantic gap between visual and textual data. Experiments demonstrate substantial improvements from using additional modalities compared to relying on images alone. Notably, our model achieves the highest reported performance on the ROUGE-L metric compared to relevant state-of-the-art models in the literature. Furthermore, we employed both human evaluation and clinical semantic similarity measurement alongside word-overlap metrics to improve the depth of quantitative analysis. A human evaluation, conducted by a board-certified radiologist, confirms the model's accuracy in identifying high-level findings, however, it also highlights that more improvement is needed to capture nuanced details and clinical context.

Emergence of number sense through the integration of multimodal information: developmental learning insights from neural network models.

Associating multimodal information is essential for human cognitive abilities including mathematical skills. Multimodal learning has also attracted attention in the field of machine learning, and it has been suggested that the acquisition of better latent representation plays an important role in enhancing task performance. This study aimed to explore the impact of multimodal learning on representation, and to understand the relationship between multimodal representation and the development of mathematical skills. We employed a multimodal deep neural network as the computational model for multimodal associations in the brain. We compared the representations of numerical information, that is, handwritten digits and images containing a variable number of geometric figures learned through single- and multimodal methods. Next, we evaluated whether these representations were beneficial for downstream arithmetic tasks. Multimodal training produced better latent representation in terms of clustering quality, which is consistent with previous findings on multimodal learning in deep neural networks. Moreover, the representations learned using multimodal information exhibited superior performance in arithmetic tasks. Our novel findings experimentally demonstrate that changes in acquired latent representations through multimodal association learning are directly related to cognitive functions, including mathematical skills. This supports the possibility that multimodal learning using deep neural network models may offer novel insights into higher cognitive functions.

An End-to-End Benchmarking Tool for Analyzing the Hardware-Software Implications of Multi-modal DNNs

Abstract-Multi-modal deep neural networks (DNNs) have become increasingly pervasive in many machine learning application domains due to their superior accuracy by fusing various modalities together. However, multi-modal DNNs present many unique characteristics such as multi-stage execution, frequent synchronization and high heterogeneity, which are not well understood in the system and architecture community. In this article, we first present and characterize a set of multi-modal DNN workloads of different sizes from five domains and measure metrics like accuracy to ensure the availability of these applications from the algorithm perspective. We then explore their important hardwaresoftware implications from system and architecture aspects by conducting an in-depth analysis on the unique hardware-software characteristics of multimodal DNNs. We hope that our work can help guide future hardware-software design and optimization for efficient inference of multi-modal DNN applications on both cloud and edge computing platforms.

Dominating Alzheimer's disease diagnosis with deep learning on sMRI and DTI-MD.

Alzheimer's disease (AD) is a progressive and irreversible neurodegenerative disorder that has become one of the major health concerns for the elderly. Computer-aided AD diagnosis can assist doctors in quickly and accurately determining patients' severity and affected regions. In this paper, we propose a method called MADNet for computer-aided AD diagnosis using multimodal datasets. The method selects ResNet-10 as the backbone network, with dual-branch parallel extraction of discriminative features for AD classification. It incorporates long-range dependencies modeling using attention scores in the decision-making layer and fuses the features based on their importance across modalities. To validate the effectiveness of our proposed multimodal classification method, we construct a multimodal dataset based on the publicly available ADNI dataset and a collected XWNI dataset, which includes examples of AD, Mild Cognitive Impairment (MCI), and Cognitively Normal (CN). On this dataset, we conduct binary classification experiments of AD vs. CN and MCI vs. CN, and demonstrate that our proposed method outperforms other traditional single-modal deep learning models. Furthermore, this conclusion also confirms the necessity of using multimodal sMRI and DTI data for computer-aided AD diagnosis, as these two modalities complement and convey information to each other. We visualize the feature maps extracted by MADNet using Grad-CAM, generating heatmaps that guide doctors' attention to important regions in patients' sMRI, which play a crucial role in the development of AD, establishing trust between human experts and machine learning models. We propose a simple yet effective multimodal deep convolutional neural network model MADNet that outperforms traditional deep learning methods that use a single-modality dataset for AD diagnosis.

Multimodal Deep Neural Networks for Digitized Document Classification

Multimodal Deep Neural Networks for Digitized Document Classification

JointNet: A Multimodal Deep Learning-Based Approach for Joint Inversion of Rayleigh Wave Dispersion and Ellipticity

ABSTRACT Joint inversion of multitype datasets is an effective approach for high-precision subsurface imaging. We present a new deep learning-based method to jointly invert Rayleigh wave phase velocity and ellipticity into shear-wave velocity of the crust and uppermost mantle. A multimodal deep neural network (termed JointNet) is designed to analyze these two independent physical parameters and generate outputs, including velocity and layer thicknesses. JointNet is trained using random 1D models and corresponding synthetic phase velocity and ellipticity, resulting in a low cost for the training dataset. Evaluation using synthetic and observed data shows that JointNet produces highly comparable results compared to those from a Markov chain Monte Carlo-based method and significantly improves inversion speed. Training using synthetic data ensures its generalized application in various regions with different velocity structures. Moreover, JointNet can be easily extended to include additional datatypes and act as a joint inversion framework to further improve imaging resolution.

Predicting multiple sclerosis severity with multimodal deep neural networks

Multiple Sclerosis (MS) is a chronic disease developed in the human brain and spinal cord, which can cause permanent damage or deterioration of the nerves. The severity of MS disease is monitored by the Expanded Disability Status Scale, composed of several functional sub-scores. Early and accurate classification of MS disease severity is critical for slowing down or preventing disease progression via applying early therapeutic intervention strategies. Recent advances in deep learning and the wide use of Electronic Health Records (EHR) create opportunities to apply data-driven and predictive modeling tools for this goal. Previous studies focusing on using single-modal machine learning and deep learning algorithms were limited in terms of prediction accuracy due to data insufficiency or model simplicity. In this paper, we proposed the idea of using patients’ multimodal longitudinal and longitudinal EHR data to predict multiple sclerosis disease severity in the future. Our contribution has two main facets. First, we describe a pioneering effort to integrate structured EHR data, neuroimaging data and clinical notes to build a multi-modal deep learning framework to predict patient’s MS severity. The proposed pipeline demonstrates up to 19% increase in terms of the area under the Area Under the Receiver Operating Characteristic curve (AUROC) compared to models using single-modal data. Second, the study also provides valuable insights regarding the amount useful signal embedded in each data modality with respect to MS disease prediction, which may improve data collection processes.

Dissolved oxygen concentration inversion based on Himawari-8 data and deep learning: a case study of lake Taihu

Dissolved Oxygen (DO) concentration is an essential water quality parameter widely used in water environments and pollution assessments, which indirectly reflects the pollution level and the occurrence of blue-green algae. With the advancement of satellite technology, the use of remote sensing techniques to estimate DO concentration has become a crucial means of water quality monitoring. In this study, we propose a novel model for DO concentration estimation in water bodies, termed Dissolved Oxygen Multimodal Deep Neural Network (DO-MDNN), which utilizes synchronous satellite remote sensing data for real-time DO concentration inversion. Using Lake Taihu as a case study, we validate the DO-MDNN model using Himawari-8 (H8) satellite imagery as input data and actual DO concentration in Lake Taihu as output data. The research results demonstrate that the DO-MDNN model exhibits high accuracy and stability in DO concentration inversion. For Lake Taihu, the performance metrics including adj_R2, RMSE, Pbias, and SMAPE are 0.77, 0.66 mg/L, −0.44%, and 5.36%, respectively. Compared to the average performance of other machine learning models, the adj_R2 shows an improvement of 6.40%, RMSE is reduced by 8.27%, and SMAPE is decreased by 12.1%. These findings highlight the operational feasibility of real-time DO concentration inversion using synchronous satellite data, providing a more efficient, economical, and accurate approach for real-time DO monitoring. This method holds significant practical value in enhancing the efficiency and precision of water environment monitoring.

Pose Graph Relocalization with Deep Object Detection and BIM-Supported Object Landmark Dictionary

Indoor localization is a prerequisite for autonomous robot applications in the construction industry. However, traditional localization techniques rely on low-level features and do not exploit construction-related semantics. They also are sensitive to environmental factors such as illumination and reflection rate, and therefore suffer from unexpected drifts and failures. This study proposes a pose graph relocalization framework that utilizes object-level landmarks to enhance a traditional visual localization system. The proposed framework builds an object landmark dictionary from Building Information Model (BIM) as prior knowledge. Then a multimodal deep neural network (DNN) is proposed to realize 3D object detection in real time, followed by instance-level object association with false-positive rejection, and relative pose estimation with outlier removal. Finally, a keyframe-based graph optimization is performed to rectify the drifts of traditional visual localization. The proposed framework was validated using a mobile platform with red-green-blue-depth (RGB-D) and inertial sensors, and the test scene was an indoor office environment with furnishing elements. The object detection model achieved 62.9% mean average precision (mAP). The relocalization technique reduced translational drifts by 64.67% and rotational drifts by 41.59% compared with traditional visual–inertial odometry.

Multimodal Deep Neural Network-Based Sensor Data Anomaly Diagnosis Method for Structural Health Monitoring

Due to sensor failure, noise interference and other factors, the data collected in the structural health monitoring (SHM) system will show a variety of abnormal patterns, which will bring great uncertainty to the structural safety assessment. This paper proposes an automatic data anomaly diagnosis method for SHM based on a multimodal deep neural network. In order to improve the detection accuracy, both two-dimensional and one-dimensional features of the sensor data are fused in the multimodal deep neural network. The network consists of two convolutional neural network (CNN) channels, one a 2D-CNN channel for extracting time–frequency features of sensor data and the other a 1D-CNN channel for extracting raw one-dimensional features of sensor data. After convolution and pooling operations for the sensor data by the 2D channel and 1D channel separately, the two types of extracted features are flattened into one-dimensional vectors and concatenated at the concatenation layer. The concatenated vector is then fed into fully connected layers for final SHM data anomaly classification. In order to evaluate the reliability of the proposed method, the monitored data lasting for one month of a long-span cable-stayed bridge were used for training, validation, and testing. Six types of training conditions (missing, minor, outlier, over-range oscillation, trend, and drift) are studied and analyzed to address the issue of imbalanced training data. With an accuracy rate of 95.10%, the optimal model demonstrates the effectiveness and capability of the proposed method. The proposed method shows a promising future as a reliable AI-assisted digital tool for safety assessment in structural health monitoring systems.

Spatial-Temporal Preserving Multimodal Algorithm for EEG-based Emotion Recognition

Emotion recognition has been the focus of some human computer interaction researchers. Designing an efficient algorithm for emotion recognition is crucial for its potential use in HCI. Many algorithms have been developed for emotion recognition and a majority of them is based on electroencephalography (EEG). Despite the success of spatial–temporal recurrent neural network (STRNN) and multimodal deep neural networks, they can still be improved. This paper proposes an algorithm which tries to combine STRNN with multimodal deep neural network, where EEG signals are classified using STRNN and eye movement signals are classified using SVM. The results from the two classifiers are fused by max fusion. This paper shows that by taking spatial and temporal information of EEG signals into consideration and using multimodal fusion, the performance of emotion recognition can be improved.

Fine-Grained 3D Modeling and Semantic Mapping of Coral Reefs Using Photogrammetric Computer Vision and Machine Learning.

Corals play a crucial role as the primary habitat-building organisms within reef ecosystems, forming expansive structures that extend over vast distances, akin to the way tall buildings define a city's skyline. However, coral reefs are vulnerable to damage and destruction due to their inherent fragility and exposure to various threats, including the impacts of climate change. Similar to successful city management, the utilization of advanced underwater videography, photogrammetric computer vision, and machine learning can facilitate precise 3D modeling and the semantic mapping of coral reefs, aiding in their careful management and conservation to ensure their survival. This study focuses on generating detailed 3D mesh models, digital surface models, and orthomosaics of coral habitats by utilizing underwater coral images and control points. Furthermore, an innovative multi-modal deep neural network is designed to perform the pixel-wise semantic segmentation of orthomosaics, enabling the projection of resulting semantic maps onto a 3D space. Notably, this study achieves a significant milestone by accomplishing semantic fine-grained 3D modeling and rugosity evaluation of coral reefs with millimeter-level accuracy, providing a potent means to understand coral reef variations under climate change with high spatial and temporal resolution.