Non-overlapping Patches Research Articles

Information sparsity guided transformer for multi-modal medical image super-resolution

Multi-modal medical image super-resolution (SR) plays a vital role in enhancing the resolution of medical images, providing more detailed visuals that aid in accurate clinical diagnosis. Recently, Transformer-based super-resolution methods have significantly promoted performance improvement in this field due to their capacity to capture global dependencies. They usually process all non-overlapping patches as tokens and densely sample these tokens without screening for calculating attention mechanisms. However, this strategy ignores the spatial sparsity of medical images, resulting in redundant or even detrimental computations on less informative regions. Hence, this paper proposes a novel sparsity-guided medical image SR network, namely SG-SRNet, by exploiting the spatial sparsity characteristics of the medical images. SG-SRNet mainly consists of two components: a sparsity mask (SM) generator for image sparsity estimation, and a sparsity-guided Transformer (SGTrans) for high-resolution image reconstruction. Specially, the SM generator generates a sparsity mask by minimizing our cross-sparsity loss which can respond to the informative positions. SGTrans first screens out the informative patches according to the sparsity mask. Then, SGTrans utilizes the designed cluster-based attention to only calculate attention between information-related tokens. We perform comprehensive experiments on three datasets to show that SG-SRNet brings significant performance enhancements with low computational complexity.

DHBLSN: A diligent hierarchical broad learning system network for cogent polyp segmentation

In medical practice, polyp segmentation holds immense significance for early Colorectal Cancer diagnosis. Over the past decade, techniques based on Deep Learning (DL) have been used extensively for segmentation purposes, showcasing promising performance. However, the efficacy of DL methods comes with a trade-off as they require manipulating many parameters for better performance, which increases the computation cost and takes a large amount of training time. In this study, we unfold a new perspective on polyp segmentation based on a broad learning system (BLS) namely dHBLSN: A diligent Hierarchical Broad Learning System Network for Cogent Polyp Segmentation that does not require multiple layers for training which significantly reduces the computational cost. The proposed dHBLSN is characterized by multi-feature extraction and hierarchical structure. Firstly, the colonoscopy images were partitioned into non-overlapping patches, which was beneficial for extracting local spatial information. Subsequently, Multi-scale features namely “Single Iteration forward pass convolution (Sifp-conv)” feature and “2D Dual Tree-Complex Wavelet Transform (2D-DT-CWT)” feature were extracted from each patch and were used directly as the feature nodes and no explicit feature mapping was required, unlike other BLS variants. These multiscale features are then enhanced and trained separately within two parallel BLS, each with its own set of feature and enhancement nodes. A separate fusion BLS was used to stack the output of the two parallel BLS hierarchically. Additionally, we introduced diligence parameter (d), to regulate the possibility of sudden divergence of the already learned and settled network during incremental learning, ensuring network stability. We performed extensive experiments on two public datasets, namely CVC-Clinic and Kvasir-SEG. Our proposed framework achieved Dice Coefficient (DC)- 0.889, Precision-0.917, Recall-0.906, F1-0.911, and F2-0.908 on the CVC-Clinic dataset. For the Kvasir-SEG dataset, our method achieved Dice Coefficient (DC)- 0.909, Precision-0.924, Recall-0.909, F1-0.916, and F2-0.911. Cross-dataset validation further underscores the generalization capability of dHBLSN, affirming its clinical relevance. Comparative analysis against state-of-the-art models highlights dHBLSN’s superior balance between effectiveness and computational complexity.

Percolation thresholds of disks with random nonoverlapping patches on four regular two-dimensional lattices.

In percolation of patchy disks on lattices, each site is occupied by a disk, and neighboring disks are regarded as connected when their patches contact. Clusters of connected disks become larger as the patchy coverage of each disk χ increases. At the percolation threshold χ_{c}, an incipient cluster begins to span the whole lattice. For systems of disks with n symmetric patches on Archimedean lattices, a recent work [Wang et al., Phys. Rev. E 105, 034118 (2022)2470-004510.1103/PhysRevE.105.034118] found symmetric properties of χ_{c}(n), which are due to the coupling of the patches' symmetry and the lattice geometry. How does χ_{c} behave with increasing n if the patches are randomly distributed on the disks? We consider two typical random distributions of the patches, i.e., the equilibrium distribution and a distribution from random sequential adsorption. Combining Monte Carlo simulations and the critical polynomial method, we numerically determine χ_{c} for 106 models of different n on the square, honeycomb, triangular, and kagome lattices. The rules governing χ_{c}(n) are investigated in detail. They are quite different from those for disks with symmetric patches and could be useful for understanding similar systems.

Coverless Image Steganography Using Content-Based Image Patch Retrieval

Abstract Image steganography is the process of concealing secret information within a cover image. The main challenge of steganography is to ensure that the embedding process does not significantly alter the cover file. In this paper, instead of modifying a cover image to carry information, steganography is performed using a set of images. These images are selected from a dataset of natural images. Each image in the dataset is divided into a number of non-overlapping patches. Then, indexing of the patches is performed based on their features. The secret image is also divided into a set of non-overlapping patches. Similar versions of the patches in the secret image are searched in the dataset to identify candidate patches. The final candidate is selected by calculating the minimum distance between the feature vector of the patches in the secret image and the patches in the dataset. Finally, the receiver retrieves the secret image using the pieces of selected images. Since, instead of embedding information in a cover image, a set of patches from natural images are selected without any changes, this approach can resist change-tracking tools, as demonstrated by experimental results, and also offers the advantage of high embedding capacity.

From patches to objects: exploiting spatial reasoning for better visual representations

AbstractAs the field of deep learning steadily transitions from the realm of academic research to practical application, the significance of self-supervised pretraining methods has become increasingly prominent. These methods, particularly in the image domain, offer a compelling strategy to effectively utilize the abundance of unlabeled image data, thereby enhancing downstream tasks’ performance. In this paper, we propose Spatial Reasoning, a novel auxiliary pretraining method that takes advantage of a more flexible formulation of contrastive learning by introducing spatial reasoning as an auxiliary task for discriminative self-supervised methods. Spatial Reasoning works by having the network predict the relative distances between sampled non-overlapping patches. We argue that this forces the network to learn more detailed and intricate internal representations of the objects and the relationships between their constituting parts. Our experiments demonstrate substantial improvement in downstream performance in linear evaluation compared to similar work and provide directions for further research into spatial reasoning.

ProDiv: Prototype-driven consistent pseudo-bag division for whole-slide image classification

Background and ObjectivePathology image classification is one of the most essential auxiliary processes in cancer diagnosis. To overcome the problem of inadequate Whole-Slide Image (WSI) samples with weak labels, pseudo-bag-based multiple instance learning (MIL) methods have attracted wide attention in pathology image classification. In this type of method, the division scheme of pseudo-bags is usually a primary factor affecting classification performance. In order to improve the division of WSI pseudo-bags on existing random/clustering approaches, this paper proposes a new Prototype-driven Division (ProDiv) scheme for the pseudo-bag-based MIL classification framework on pathology images. MethodsThis scheme first designs an attention-based method to generate a bag prototype for each slide. On this basis, it further groups WSI patch instances into a series of instance clusters according to the feature similarities between the prototype and patches. Finally, pseudo-bags are obtained by randomly combining the non-overlapping patch instances of different instance clusters. Moreover, the design scheme of our ProDiv considers practicality, and it could be smoothly assembled with almost all the MIL-based WSI classification methods in recent years. ResultsEmpirical results show that our ProDiv, when integrated with several existing methods, can deliver classification AUC improvements of up to 7.3% and 10.3%, respectively on two public WSI datasets. ConclusionsProDiv could almost always bring obvious performance improvements to compared MIL models on typical metrics, which suggests the effectiveness of our scheme. Experimental visualization also visually interprets the correctness of the proposed ProDiv.

Patched Line Segment Learning for Vector Road Mapping

This paper presents a novel approach to computing vector road maps from satellite remotely sensed images, building upon a well-defined Patched Line Segment (PaLiS) representation for road graphs that holds geometric significance. Unlike prevailing methods that derive road vector representations from satellite images using binary masks or keypoints, our method employs line segments. These segments not only convey road locations but also capture their orientations, making them a robust choice for representation. More precisely, given an input image, we divide it into non-overlapping patches and predict a suitable line segment within each patch. This strategy enables us to capture spatial and structural cues from these patch-based line segments, simplifying the process of constructing the road network graph without the necessity of additional neural networks for connectivity. In our experiments, we demonstrate how an effective representation of a road graph significantly enhances the performance of vector road mapping on established benchmarks, without requiring extensive modifications to the neural network architecture. Furthermore, our method achieves state-of-the-art performance with just 6 GPU hours of training, leading to a substantial 32-fold reduction in training costs in terms of GPU hours.

Seizure prediction based on improved vision transformer model for EEG channel optimization

Epileptic seizures are unpredictable events caused by abnormal discharges of a patient’s brain cells. Extensive research has been conducted to develop seizure prediction algorithms based on long-term continuous electroencephalogram (EEG) signals. This paper describes a patient-specific seizure prediction method that can serve as a basis for the design of lightweight, wearable and effective seizure-prediction devices. We aim to achieve two objectives using this method. The first aim is to extract robust feature representations from multichannel EEG signals, and the second aim is to reduce the number of channels used for prediction by selecting an optimal set of channels from multichannel EEG signals while ensuring good prediction performance. We design a seizure-prediction algorithm based on a vision transformer (ViT) model. The algorithm selects channels that play a key role in seizure prediction from 22 channels of EEG signals. First, we perform a time-frequency analysis of processed time-series signals to obtain EEG spectrograms. We then segment the spectrograms of multiple channels into many non-overlapping patches of the same size, which are input into the channel selection layer of the proposed model, named Sel-JPM-ViT, enabling it to select channels. Application of the Sel-JPM-ViT model to the Boston Children’s Hospital–Massachusetts Institute of Technology scalp EEG dataset yields results using only three to six channels of EEG signals that are slightly better that the results obtained using 22 channels of EEG signals. Overall, the Sel-JPM-ViT model exhibits an average classification accuracy of 93.65%, an average sensitivity of 94.70% and an average specificity of 92.78%.

Lymph node metastases detection in Whole Slide Images using prototypical patterns and transformer-guided multiple instance learning

Abstract Background: The examination of lymph nodes (LNs) regarding metastases is vital for the staging of cancer patients, which is necessary for diagnosis and adequate treatment selection. Advancements in digital pathology, utilizing Whole-Slide Images (WSIs) and convolutional neural networks (CNNs), pose new opportunities to automate this procedure, thus reducing pathologists’ workload while simultaneously increasing the accuracy in metastases detection. Objective: To address the task of LN-metastases detection, the use of weakly supervised transformers are applied for the analysis of WSIs. Methods & Materials: As WSIs are too large to be processed as a whole, they are divided into non-overlapping patches, which are converted to feature vectors using a CNN network, pre-trained on HE-stained colon cancer resections. A subset of these patches serves as input for a transformer to predict if a LN contains a metastasis. Hence, selecting a representative subset is an important part of the pipeline. Hereby, a prototype based clustering is employed and different sampling strategies are tested. Finally, the chosen feature vectors are fed into a transformer-based multiple instance learning (MIL) architecture, classifying the LNs into healthy/negative (that is, containing no metastases), or metastatic/positive (that is, containing metastases). The proposed model is trained only on the Camelyon16 training data (LNs from breast cancer patients), and evaluated on the Camelyon16 test set. Results: The trained model achieves accuracies of up to 92.3% on the test data (from breast LNs). While the model struggles with smaller metastases, high specificities of up to 96.9% can be accomplished. Additionally, the model is evaluated on LNs from a different primary tumor (colon), where accuracies between 62.3% and 95.9% could be obtained. Conclusion: The investigated transformer-model performs very good on LN data from the public LN breast data, but the domain transfer to LNs from the colon needs more research.

SSAPN: Spectral–Spatial Anomaly Perception Network for Unsupervised Vaccine Detection

Vaccines are the most significant and effective way to prevent disease and safeguard human health. However, it is easy to produce or mix foreign matters during the manufacturing process. Moreover, foreign matters are extremely faint that it is difficult to obtain images and detect them accurately. To tackle imaging challenges, in this paper, we built a hyperspectral imaging system to construct a first-of-its-kind HSI dataset with pixel-level annotation for vaccine anomaly detection, where the vaccine comes from the actual pharmaceutical company. To address the problem of low detection accuracy, we propose a spectral-spatial anomaly perception module joint with an unsupervised Auto-Encoder network (SSAPN), in which nonlinear features learned from the encoder are divided into non-overlapping patches and mapped to efficiently encode spectral and spatial feature information. The spectral-spatial MLP module consists of continuous and alternating spectral MLP with spatial MLP, which achieves spectral with spatial perception in the global receptive field, captures long-range dependencies, and extracts the most discriminative spectral-spatial features. Experimental results show that our SSAPN model outperforms other state-of-the-art anomaly detection methods in terms of both detection and generalization performance. This work will help speed up the production process in the vaccine pharmaceutical industry and ensure vaccine quality.

ShuffleDetect: Detecting Adversarial Images against Convolutional Neural Networks

Recently, convolutional neural networks (CNNs) have become the main drivers in many image recognition applications. However, they are vulnerable to adversarial attacks, which can lead to disastrous consequences. This paper introduces ShuffleDetect as a new and efficient unsupervised method for the detection of adversarial images against trained convolutional neural networks. Its main feature is to split an input image into non-overlapping patches, then swap the patches according to permutations, and count the number of permutations for which the CNN classifies the unshuffled input image and the shuffled image into different categories. The image is declared adversarial if and only if the proportion of such permutations exceeds a certain threshold value. A series of 8 targeted or untargeted attacks was applied on 10 diverse and state-of-the-art ImageNet-trained CNNs, leading to 9500 relevant clean and adversarial images. We assessed the performance of ShuffleDetect intrinsically and compared it with another detector. Experiments show that ShuffleDetect is an easy-to-implement, very fast, and near memory-free detector that achieves high detection rates and low false positive rates.

Assessing skin blood flow function in people with spinal cord injury using the time domain, time–frequency domain and deep learning approaches

Skin blood flow (SBF) has been assessed using the time domain and time–frequency domain methods. However, these methods require prior knowledge of selecting appropriate parameters for characterizing SBF responses. Deep learning has been successful on classification of medical images, and could be a promising tool for assessing SBF in various pathophysiological conditions. In this study, we proposed a deep learning-based framework for converting 1-dimensional time-series SBF into 2-dimensional time–frequency SBF for convolutional neural networks (CNNs). Thirty-seven participants were recruited into this study, including 21 people with spinal cord injury (SCI) and 16 healthy able-bodied controls. Laser Doppler flowmetry was used to measure sacral SBF. Continuous wavelet transform was used to obtain time–frequency representations of SBF. The whole frequency (WF, 0.0095–2 Hz), high frequency (HF, 0.138–2 Hz), and low frequency (LF, 0.0095–0.138 Hz) regions of the wavelet amplitudes were partitioned into the nonoverlapping patches. Four CNNs including AlexNet, Vgg-19, GoogLeNet, and ResNet-18 were employed to classify the patches. The results showed that the time-domain biphasic thermal index could not differentiate SBF in all groups. Time-frequency wavelet analysis showed differences in myogenic and cardiac controls between people with SCI who were active and sedentary (p < 0.01). CNNs results showed that all participants could be correctly classified based on the WF patches (100% of accuracy) compared to the HF (50–100%) and LF (66.7–100%) patches and five individual oscillation components (50–57.1%). Our study demonstrated that the classifiers could detect subtle changes in SBF function that cannot be revealed by the traditional methods.

Discrete Cosine Transformed Images Are Easy to Recognize in Vision Transformers

Deep learning models for image classification with adequate parameters show excellent classification performance because they can effectively extract the features of input images. On the other hand, there is a limit to the abilities of deep learning models to interpret images using only spatial information because an image is a signal with great spatial redundancy. Therefore, in this study, the discrete cosine transform was applied to an input image in units of an N×N block size to allow the deep learning model to employ both frequency and spatial information. The proposed method was implemented and verified by selecting a vision transformer using a 16×16 nonoverlapping patch as a baseline and training various datasets of Cifar-10, Cifar-100, and Tiny-ImageNet from the very beginning without pre-trained weights. The experimental results showed that the top-1 accuracy is improved by approximately 3-5% for every dataset with little increase in computational cost.

Image Inpainting with Bilateral Convolution

Due to sensor malfunctions and poor atmospheric conditions, remote sensing images often miss important information/pixels, which affects downstream tasks, therefore requiring reconstruction. Current image reconstruction methods use deep convolutional neural networks to improve inpainting performances as they have a powerful modeling capability. However, deep convolutional networks learn different features with the same group of convolutional kernels, which restricts their ability to handle diverse image corruptions and often results in color discrepancy and blurriness in the recovered images. To mitigate this problem, in this paper, we propose an operator called Bilateral Convolution (BC) to adaptively preserve and propagate information from known regions to missing data regions. On the basis of vanilla convolution, the BC dynamically propagates more confident features, which weights the input features of a patch according to their spatial location and feature value. Furthermore, to capture different range dependencies, we designed a Multi-range Window Attention (MWA) module, in which the input feature is divided into multiple sizes of non-overlapped patches for several heads, and then these feature patches are processed by the window self-attention. With BC and MWA, we designed a bilateral convolution network for image inpainting. We conducted experiments on remote sensing datasets and several typical image inpainting datasets to verify the effectiveness and generalization of our network. The results show that our network adaptively captures features between known and unknown regions, generates appropriate content for various corrupted images, and has a competitive performance compared with state-of-the-art methods.

Efficient and Adaptable Patch-Based Crack Detection

Road crack inspection is an important process to maintain the quality of roads for safety issues. The manual road inspection is laborious and time consuming. Thus, an automatic road crack detection is essential to make the inspection process easier and faster. Normally, this crack detection is conducted in real-time with low power computational devices and limited memory. Consequently, many former approaches adopted the patch-based approach to reduce computation for single forward, where the input image is divided into several non-overlapping patches. The generated patches are processed by independent CNN models to capture the crack position. Yet, this approach can lead to disintegrating issues because each patch is processed independently when the CNN fails to detect some patches of the crack. In this study, the improved patch-based crack detector is proposed, and the global patch analyzer is adopted to handle the above issue by considering the relation of each patch processing. Moreover, the proposed model also involves more features such as multiple decoder design and automatic resource mapper to yield superior results than that of the state-of-the-art methods in terms of the speed and accuracy as examined by extensive experiments.

Brain Tumor Classification Using Conditional Segmentation with Residual Network and Attention Approach by Extreme Gradient Boost

A brain tumor is a tumor in the brain that has grown out of control, which is a dangerous condition for the human body. For later prognosis and treatment planning, the accurate segmentation and categorization of cancers are crucial. Radiologists must use an automated approach to identify brain tumors, since it is an error-prone and time-consuming operation. This work proposes conditional deep learning for brain tumor segmentation, residual network-based classification, and overall survival prediction using structural multimodal magnetic resonance images (MRI). First, we propose conditional random field and convolution network-based segmentation, which identifies non-overlapped patches. These patches need minimal time to identify the tumor. If they overlap, the errors increase. The second part of this paper proposes residual network-based feature mapping with XG-Boost-based learning. In the second part, the main emphasis is on feature mapping in nonlinear space with residual features, since residual features reduce the chances of loss information, and nonlinear space mapping provides efficient tumor information. Features mapping learned by XG-Boost improves the structural-based learning and increases the accuracy class-wise. The experiment uses two datasets: one for two classes (cancer and non-cancer) and the other for three classes (meningioma, glioma, pituitary). The performance on both improves significantly compared to another existing approach. The main objective of this research work is to improve segmentation and its impact on classification performance parameters. It improves by conditional random field and residual network. As a result, two-class accuracy improves by 3.4% and three-class accuracy improves by 2.3%. It is enhanced with a small convolution network. So, we conclude in fewer resources, and better segmentation improves the results of brain tumor classification.

3D time-dependent scattering about complex shapes using high order difference potentials

We compute the scattering of unsteady acoustic waves about complex three-dimensional bodies with high order accuracy. The geometry of a scattering body is defined with the help of CAD. Its surface is represented as a collection of non-overlapping patches, each parameterized independently by means of high order splines (NURBS). As a specific example, we consider a submarine-like scatterer constructed using five different patches.The acoustic wave equation on the region exterior to the scatterer is solved by first reducing it to a system of Calderon's boundary operator equations. The latter are obtained using the method of difference potentials coupled with a compact fourth order accurate finite difference scheme. When solving the boundary operator equations, we employ Huygens' principle. It allows us to work on a sliding time window of non-increasing duration rather than keep the entire temporal history of the solution at the boundary.The proposed methodology demonstrates grid-independent computational complexity at the boundary and sub-linear complexity with respect to the grid dimension. It efficiently handles complex non-conforming geometries on Cartesian grids with no penalty for either accuracy or stability due to the cut cells. Its performance does not deteriorate over arbitrarily long simulation times. The exact treatment of artificial outer boundaries is inherently built in. Finally, multiple similar problems can be solved efficiently at a low individual cost per problem. This is important when, for example, the boundary condition on the surface changes but the scattering body stays the same.

Virtual CT Myelography: A Patch-Based Machine Learning Model to Improve Intraspinal Soft Tissue Visualization on Unenhanced Dual-Energy Lumbar Spine CT

Background: Distinguishing between the spinal cord and cerebrospinal fluid (CSF) non-invasively on CT is challenging due to their similar mass densities. We hypothesize that patch-based machine learning applied to dual-energy CT can accurately distinguish CSF from neural or other tissues based on the center voxel and neighboring voxels. Methods: 88 regions of interest (ROIs) from 12 patients’ dual-energy (100 and 140 kVp) lumbar spine CT exams were manually labeled by a neuroradiologist as one of 4 major tissue types (water, fat, bone, and nonspecific soft tissue). Four-class classifier convolutional neural networks were trained, validated, and tested on thousands of nonoverlapping patches extracted from 82 ROIs among 11 CT exams, with each patch representing pixel values (at low and high energies) of small, rectangular, 3D CT volumes. Different patch sizes were evaluated, ranging from 3 × 3 × 3 × 2 to 7 × 7 × 7 × 2. A final ensemble model incorporating all patch sizes was tested on patches extracted from six ROIs in a holdout patient. Results: Individual models showed overall test accuracies ranging from 99.8% for 3 × 3 × 3 × 2 patches (N = 19,423) to 98.1% for 7 × 7 × 7 × 2 patches (N = 1298). The final ensemble model showed 99.4% test classification accuracy, with sensitivities and specificities of 90% and 99.6%, respectively, for the water class and 98.6% and 100% for the soft tissue class. Conclusions: Convolutional neural networks utilizing local low-level features on dual-energy spine CT can yield accurate tissue classification and enhance the visualization of intraspinal neural tissue.

Deep Learning Feature Extraction Using Attention-Based DenseNet 121 for Copy Move Forgery Detection

In modern society, digital images can be far-reaching, and the images are manipulated by various software and hardware technologies. The image forgery activities are undertaken by the attackers mainly for damaging the reputation of people or receiving fiscal gain, etc. Taking this into consideration, many techniques are developed to detect the forged images. In this paper, a new deep learning-based approach is introduced for copy-move forgery detection. The input images are segmented into non-overlapping patches using superpixel-based modified dense peak clustering and the features are extracted from the segmented patches by applying deep learning structure of attention-based DenseNet 121 model. Besides, to compare every block, the depth of each pixel is reconstructed, and eventually matching process is carried out using the adaptive chimp patch matching approach, which detects the suspicious forged regions in an image. Finally, the matched keypoints are merged with the segmented patches using the merged keypoint matching algorithm. As a result, the new deep learning approach has detected the forged regions efficiently from the tampered image with less time even the image is compressed, rotated, or scaled. The performance is evaluated in terms of recall, precision, accuracy, F1-score, computational time, and False Positive Rate (FPR). Moreover, the performance is compared with the other existing approaches, and the outcomes showed that the proposed method has achieved higher accuracy of 97%, recall of 99%, precision of 97.84%, F1-score of 98.81%, FPR of 0.022 and less computational time of 2.5 s.

Multi-Channel Vision Transformer for Epileptic Seizure Prediction.

Epilepsy is a neurological disorder that causes recurrent seizures and sometimes loss of awareness. Around 30% of epileptic patients continue to have seizures despite taking anti-seizure medication. The ability to predict the future occurrence of seizures would enable the patients to take precautions against probable injuries and administer timely treatment to abort or control impending seizures. In this study, we introduce a Transformer-based approach called Multi-channel Vision Transformer (MViT) for automated and simultaneous learning of the spatio-temporal-spectral features in multi-channel EEG data. Continuous wavelet transform, a simple yet efficient pre-processing approach, is first used for turning the time-series EEG signals into image-like time-frequency representations named Scalograms. Each scalogram is split into a sequence of fixed-size non-overlapping patches, which are then fed as inputs to the MViT for EEG classification. Extensive experiments on three benchmark EEG datasets demonstrate the superiority of the proposed MViT algorithm over the state-of-the-art seizure prediction methods, achieving an average prediction sensitivity of 99.80% for surface EEG and 90.28–91.15% for invasive EEG data.