Decoder Network Research Articles

A Convolution Auto-Encoders Network for Aero-Engine Hot Jet FT-IR Spectrum Feature Extraction and Classification

Aiming at classification and recognition of aero-engines, two telemetry Fourier transform infrared (FT-IR) spectrometers are utilized to measure the infrared spectrum of the areo-engine hot jet, meanwhile a spectrum dataset of six types of areo-engines is established. In this paper, a convolutional autoencoder (CAE) is designed for spectral feature extraction and classification, which is composed of coding network, decoding network, and classification network. The encoder network consists of convolutional layers and maximum pooling layers, the decoder network consists of up-sampling layers and deconvolution layers, and the classification network consists of a flattened layer and a dense layer. In the experiment, data for the spectral dataset were randomly sampled at a ratio of 8:1:1 to produce the training set, validation set, and prediction set, and the performance measures were accuracy, precision, recall, confusion matrix, F1 score, ROC curve, and AUC value. The experimental result of CAE reached 96% accuracy and the prediction running time was 1.57 s. Compared with the classical PCA feature extraction and SVM, XGBoost, AdaBoost, and Random Forest classifier algorithms, as well as AE, CSAE, and CVAE deep learning classification methods, the CAE network can achieve higher accuracy and efficiency and can complete the spectral classification task.

Polarimetric synthetic aperture radar image classification for accurate landcover mapping with advanced encoder–decoder networks using multi-dimensional probabilistic voting ensemble

Polarimetric synthetic aperture radar image classification for accurate landcover mapping with advanced encoder–decoder networks using multi-dimensional probabilistic voting ensemble

REDalign: accurate RNA structural alignment using residual encoder-decoder network

BackgroundRNA secondary structural alignment serves as a foundational procedure in identifying conserved structural motifs among RNA sequences, crucially advancing our understanding of novel RNAs via comparative genomic analysis. While various computational strategies for RNA structural alignment exist, they often come with high computational complexity. Specifically, when addressing a set of RNAs with unknown structures, the task of simultaneously predicting their consensus secondary structure and determining the optimal sequence alignment requires an overwhelming computational effort of O(L6)\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$O(L^6)$$\\end{document} for each RNA pair. Such an extremely high computational complexity makes these methods impractical for large-scale analysis despite their accurate alignment capabilities.ResultsIn this paper, we introduce REDalign, an innovative approach based on deep learning for RNA secondary structural alignment. By utilizing a residual encoder-decoder network, REDalign can efficiently capture consensus structures and optimize structural alignments. In this learning model, the encoder network leverages a hierarchical pyramid to assimilate high-level structural features. Concurrently, the decoder network, enhanced with residual skip connections, integrates multi-level encoded features to learn detailed feature hierarchies with fewer parameter sets. REDalign significantly reduces computational complexity compared to Sankoff-style algorithms and effectively handles non-nested structures, including pseudoknots, which are challenging for traditional alignment methods. Extensive evaluations demonstrate that REDalign provides superior accuracy and substantial computational efficiency.ConclusionREDalign presents a significant advancement in RNA secondary structural alignment, balancing high alignment accuracy with lower computational demands. Its ability to handle complex RNA structures, including pseudoknots, makes it an effective tool for large-scale RNA analysis, with potential implications for accelerating discoveries in RNA research and comparative genomics.

Channel reconstruction through improvised deep learning architecture for high-speed networks

Efficient acquisition of channel state information (CSI) is quite complicated process but immensely essential to exploit probable benefits of massive multiple input multiple output (MIMO) systems. Therefore, a deep learningbased model is utilized to estimate channel feedback in a massive MIMO system. The proposed improvised deep learning-based channel estimation (IDLCE) model enhances channel reconstruction efficiency by using multiple convolutional layers and residual blocks. The proposed IDLCE model utilizes encoder network to compress CSI matrices where decoder network is used to downlink reconstruct CSI matrices. Here, an additional quantization block is incorporated to improve feedback reconstruction accuracy by reducing channel errors. A COST 2,100 model is adopted to analyse performance efficiency for both indoor and outdoor scenarios. Further, deep learning-based model is used to train thousands of parameter and correlation coefficients much faster and to minimize computational complexity. The proposed IDLCE model evaluate performance in terms of normalized mean square error (NMSE), correlation efficiency and reconstruction accuracy and compared against varied state-of-art-channel estimation techniques. Excellent performance results are obtained with large improvement in channel reconstruction accuracy

DSIFNet: Implicit Feature Network for Nasal Cavity and Vestibule Segmentation from 3D Head CT

This study is dedicated to accurately segment the nasal cavity and its intricate internal anatomy from head CT images, which is critical for understanding nasal physiology, diagnosing diseases, and planning surgeries. Nasal cavity and it’s anatomical structures such as the sinuses, and vestibule exhibit significant scale differences, with complex shapes and variable microstructures. These features require the segmentation method to have strong cross-scale feature extraction capabilities. To effectively address this challenge, we propose an image segmentation network named the Deeply Supervised Implicit Feature Network (DSIFNet). This network uniquely incorporates an Implicit Feature Function Module Guided by Local and Global Positional Information (LGPI-IFF), enabling effective fusion of features across scales and enhancing the network's ability to recognize details and overall structures. Additionally, we introduce a deep supervision mechanism based on implicit feature functions in the network's decoding phase, optimizing the utilization of multi-scale feature information, thus improving segmentation precision and detail representation. Furthermore, we constructed a dataset comprising 7116 CT volumes (including 1,292,508 slices) and implemented PixPro-based self-supervised pretraining to utilize unlabeled data for enhanced feature extraction. Our tests on nasal cavity and vestibule segmentation, conducted on a dataset comprising 128 head CT volumes (including 34006 slices), demonstrate the robustness and superior performance of proposed method, achieving leading results across multiple segmentation metrics.

ST-SHAP: A hierarchical and explainable attention network for emotional EEG representation learning and decoding

Background:Emotion recognition using electroencephalogram (EEG) has become a research hotspot in the field of human–computer interaction, how to sufficiently learn complex spatial–temporal representations of emotional EEG data and obtain explainable model prediction results are still great challenges. New methodIn this study, a novel hierarchical and explainable attention network ST-SHAP which combines the Swin Transformer (ST) and SHapley Additive exPlanations (SHAP) technique is proposed for automatic emotional EEG classification. Firstly, a 3D spatial–temporal feature of emotional EEG data is generated via frequency band filtering, temporal segmentation, spatial mapping, and interpolation to fully preserve important spatial–temporal-frequency characteristics. Secondly, a hierarchical attention network is devised to sufficiently learn an abstract spatial–temporal representation of emotional EEG and perform classification. Concretely, in this decoding model, the W-MSA module is used for modeling correlations within local windows, the SW-MSA module allows for information interactions between different local windows, and the patch merging module further facilitates local-to-global multiscale modeling. Finally, the SHAP method is utilized to discover important brain regions for emotion processing and improve the explainability of the Swin Transformer model. Results:Two benchmark datasets, namely SEED and DREAMER, are used for classification performance evaluation. In the subject-dependent experiments, for SEED dataset, ST-SHAP achieves an average accuracy of 97.18%, while for DREAMER dataset, the average accuracy is 96.06% and 95.98% on arousal and valance dimension respectively. In addition, import brain regions that conform to prior knowledge of neurophysiology are discovered via a data-driven approach for both datasets. Comparison with existing methods:In terms of subject-dependent and subject-independent emotional EEG decoding accuracies, our method outperforms several closely related existing methods. Conclusion:These experimental results fully prove the effectiveness and superiority of our proposed algorithm.

Research on Land Use and Land Cover Information Extraction Methods for Remote Sensing Images Based on Improved Convolutional Neural Networks

To address the challenges that convolutional neural networks (CNNs) face in extracting small objects and handling class imbalance in remote sensing imagery, this paper proposes a novel spatial contextual information and multiscale feature fusion encoding–decoding network, SCIMF-Net. Firstly, SCIMF-Net employs an improved ResNeXt-101 deep backbone network, significantly enhancing the extraction capability of small object features. Next, a novel PMFF module is designed to effectively promote the fusion of features at different scales, deepening the model’s understanding of global and local spatial contextual information. Finally, introducing a weighted joint loss function improves the SCIMF-Net model’s performance in extracting LULC information under class imbalance conditions. Experimental results show that compared to other CNNs such as Res-FCN, U-Net, SE-U-Net, and U-Net++, SCIMF-Net improves PA by 0.68%, 0.54%, 1.61%, and 3.39%, respectively; MPA by 2.96%, 4.51%, 2.37%, and 3.45%, respectively; and MIOU by 3.27%, 4.89%, 4.2%, and 5.68%, respectively. Detailed comparisons of locally visualized LULC information extraction results indicate that SCIMF-Net can accurately extract information from imbalanced classes and small objects.

Efficient Metal Corrosion Area Detection Model Combining Convolution and Transformer

In the context of rapid industrialization, efficiently detecting metal corrosion areas has become a critical task in preventing material damage. Unlike conventional semantic segmentation targets, metal corrosion characteristics vary significantly in color, texture, and size. Traditional image segmentation methods need improvement in scenarios involving occlusions, shadows, and defects. This paper proposes a convolution and sequence encoding combined network, MCD-Net, for metal corrosion area segmentation. First, a visual Transformer sequence encoder is introduced into the convolutional encoder–decoder network to enhance global information processing capabilities and establish long-range feature dependencies. A feature fusion method based on an attention module is proposed to enhance the model’s ability to recognize corrosion boundaries, thereby enhancing segmentation accuracy and model robustness. Finally, in the model’s decoding stage, a score-based multi-scale feature enhancement method is employed to emphasize significant features in the corrosion areas. Experimental results indicate that this method attained an F1 score of 84.53% on a public corrosion dataset, demonstrating the model’s deeper understanding and reasoning capabilities for shadow and defect features, as well as excellent noise resistance performance.

RPDNet: A reconstruction-regularized parallel decoders network for rectal tumor and rectum co-segmentation

Accurate segmentation of rectal cancer tumor and rectum in magnetic resonance imaging (MRI) is significant for tumor precise diagnosis and treatment plans determination. Variable shapes and unclear boundaries of rectal tumors make this task particularly challenging. Only a few studies have explored deep learning networks in rectal tumor segmentation, which mainly adopt the classical encoder-decoder structure. The frequent downsampling operations during feature extraction result in the loss of detailed information, limiting the network's ability to precisely capture the shape and boundary of rectal tumors. This paper proposes a Reconstruction-regularized Parallel Decoder network (RPDNet) to address the problem of information loss and obtain accurate co-segmentation results of both rectal tumor and rectum. RPDNet initially establishes a shared encoder and parallel decoders framework to fully utilize the common knowledge between two segmentation labels while reducing the number of network parameters. An auxiliary reconstruction branch is subsequently introduced by calculating the consistency loss between the reconstructed and input images to preserve sufficient anatomical structure information. Moreover, a non-parameter target-adaptive attention module is proposed to distinguish the unclear boundary by enhancing the feature-level contrast between rectal tumors and normal tissues. The experimental results indicate that the proposed method outperforms state-of-the-art approaches in rectal tumor and rectum segmentation tasks, with Dice coefficients of 84.91 % and 90.36 %, respectively, demonstrating its potential application value in clinical practice.

Fault3DNnet: A lightweight 3D seismic fault detection network with bidirectional decoding

Fault detection from seismic data is a critical component of hydrocarbon exploration and production. In recent years, deep-learning techniques have made tremendous achievements and breakthroughs in seismic fault detection. Especially U-nets, represented by FaultSeg3D and its variants, are a dominant method for 3D deep-learning-based seismic fault interpretation. The incorporation of state-of-the-art network modules or training strategies with U-nets continuously improves the performance of fault detection for field seismic data, usually at the cost of increased computational intensity and hardware consumption. We design a lightweight bidirectional decoding network, Fault3DNnet, for fault detection from 3D seismic data volume. To make it simple and concise, our Fault3DNnet is equipped with basic conventional convolution and regular operation modules. The two decoding branches in Fault3DNnet could provide complementary information and improve fault detection performance. The superiority of Fault3DNnet is verified using publicly available synthetic and four field seismic data sets. The experimental results indicate that the method predicts fault structures with improved completeness and continuity compared with FaultSeg3D. Meanwhile, the parameter and computational quantity of the network are only 1/7 and 1/5 of FaultSeg3D, respectively. With the same graphics processing unit memory consumption, Fault3DNnet can handle the size of approximately six times the volume of 3D seismic data during inferencing as that of FaultSeg3D. It helps to improve prediction stability and decrease discontinuous artifacts at the stitching boundaries when predicting the large field data in chunks. In general, our Fault3DNnet delivers superior fault detection results while greatly reducing computational intensity and memory usage.

A Multi-Hierarchical Complementary Feature Interaction Network for Accelerated Multi-Modal MR Imaging

Magnetic resonance (MR) imaging is widely used in the clinical field due to its non-invasiveness, but the long scanning time is still a bottleneck for its popularization. Using the complementary information between multi-modal imaging to accelerate imaging provides a novel and effective MR fast imaging solution. However, previous technologies mostly use simple fusion methods and fail to fully utilize their potential sharable knowledge. In this study, we introduced a novel multi-hierarchical complementary feature interaction network (MHCFIN) to realize joint reconstruction of multi-modal MR images with undersampled data and thus accelerate multi-modal imaging. Firstly, multiple attention mechanisms are integrated with a dual-branch encoder–decoder network to represent shared features and complementary features of different modalities. In the decoding stage, the multi-modal feature interaction module (MMFIM) acts as a bridge between the two branches, realizing complementary knowledge transfer between different modalities through cross-level fusion. The single-modal feature fusion module (SMFFM) carries out multi-scale feature representation and optimization of the single modality, preserving better anatomical details. Extensive experiments are conducted under different sampling patterns and acceleration factors. The results show that this proposed method achieves obvious improvement compared with existing state-of-the-art reconstruction methods in both visual quality and quantity.

Remote Sensing LiDAR and Hyperspectral Classification with Multi-Scale Graph Encoder–Decoder Network

The rapid development of sensor technology has made multi-modal remote sensing data valuable for land cover classification due to its diverse and complementary information. Many feature extraction methods for multi-modal data, combining light detection and ranging (LiDAR) and hyperspectral imaging (HSI), have recognized the importance of incorporating multiple spatial scales. However, effectively capturing both long-range global correlations and short-range local features simultaneously on different scales remains a challenge, particularly in large-scale, complex ground scenes. To address this limitation, we propose a multi-scale graph encoder–decoder network (MGEN) for multi-modal data classification. The MGEN adopts a graph model that maintains global sample correlations to fuse multi-scale features, enabling simultaneous extraction of local and global information. The graph encoder maps multi-modal data from different scales to the graph space and completes feature extraction in the graph space. The graph decoder maps the features of multiple scales back to the original data space and completes multi-scale feature fusion and classification. Experimental results on three HSI-LiDAR datasets demonstrate that the proposed MGEN achieves considerable classification accuracies and outperforms state-of-the-art methods.

Invisible backdoor attack with attention and steganography

Recently, with the development and widespread application of deep neural networks (DNNs), backdoor attacks have posed new security threats to the training process of DNNs. Backdoor attacks on neural networks undermine the security and trustworthiness of DNNs by implanting hidden, unauthorized triggers, leading to benign behavior on clean samples while exhibiting malicious behavior on samples containing backdoor triggers. Existing backdoor attacks typically employ triggers that are sample-agnostic and identical for each sample, resulting in poisoned images that lack naturalness and are ineffective against existing backdoor defenses. To address these issues, this paper proposes a novel stealthy backdoor attack, where the backdoor trigger is dynamic and specific to each sample. Specifically, we leverage spatial attention on images and pre-trained models to obtain dynamic triggers, which are then injected using an encoder–decoder network. The design of the injection network benefits from recent advances in steganography research. To demonstrate the effectiveness of the proposed steganographic network, we design two backdoor attack modes named ASBA and ATBA, where ASBA utilizes the steganographic network for attack, while ATBA is a backdoor attack without steganography. Subsequently, we conducted attacks on Deep Neural Networks (DNNs) using four standard datasets. Our extensive experiments show that ASBA surpasses ATBA in terms of stealthiness and resilience against current defensive measures. Furthermore, both ASBA and ATBA demonstrate superior attack efficiency.

CEDNet: A cascade encoder–decoder network for dense prediction

The prevailing methods for dense prediction tasks typically utilize a heavy classification backbone to extract multi-scale features and then fuse these features using a lightweight module. However, these methods allocate most computational resources to the classification backbone, which delays the multi-scale feature fusion and potentially leads to inadequate feature fusion. Although some methods perform feature fusion from early stages, they either fail to fully leverage high-level features to guide low-level feature learning or have complex structures, resulting in sub-optimal performance. We propose a streamlined cascade encoder–decoder network, named CEDNet, tailored for dense prediction tasks. All stages in CEDNet share the same encoder–decoder structure and perform multi-scale feature fusion within each decoder, thereby enhancing the effectiveness of multi-scale feature fusion. We explored three well-known encoder–decoder structures: Hourglass, UNet, and FPN, all of which yielded promising results. Experiments on various dense prediction tasks demonstrated the effectiveness of our method.11Code: https://github.com/zhanggang001/CEDNet.

A Bi‐level stacked LSTM‐DNN‐based decoder network for AGC dispatch under regulation market framework in presence of VPP and EV aggregators

AbstractThe consideration of mileage settlement in the frequency regulation market has encouraged fast‐acting units, such as converter‐interfaced generators (CIG) and electric vehicle stations, to actively participate in load‐generation balancing through automatic generation control (AGC). Conventional frequency regulation faces challenges in coping with the growing variability of CIGs and also lacks effective incentives for rapid‐responding units. In this context, a bi‐level AGC dispatch approach based on a stacked long short‐term memory (LSTM)‐deep neural network (DNN)‐based decoder framework is proposed for a power system comprising diverse CIGs forming a virtual power plant and electric vehicle aggregators. The proposed decoder network is comprised of stacked LSTM and DNN, wherein the cascaded LSTM layers are introduced to accurately capture temporal information from time series input. The inclusion of a dropout mechanism further enhances the model’s generalisability in unforeseen environments. The proposed dispatch framework uses mileage‐based compensation criteria to optimally allocate instructions among various participating units with differing regulation characteristics. The performance of the proposed method is analysed by considering packet loss, delay, unexpected generation failure, and denial of service attacks. The evaluation of the proposed approach reveals its superior performance compared to proportionality, particle swarm optimisation, decision tree, and DNN methods.

Leveraging advanced AI algorithms with transformer-infused recurrent neural networks to optimize solar irradiance forecasting

Solar energy (SE) is vital for renewable energy generation, but its natural fluctuations present difficulties in maintaining grid stability and planning. Accurate forecasting of solar irradiance (SI) is essential to address these challenges. The current research presents an innovative forecasting approach named as Transformer-Infused Recurrent Neural Network (TIR) model. This model integrates a Bi-Directional Long Short-Term Memory (BiLSTM) network for encoding and a Gated Recurrent Unit (GRU) network for decoding, incorporating attention mechanisms and positional encoding. This model is proposed to enhance SI forecasting accuracy by effectively utilizing meteorological weather data, handling overfitting, and managing data outliers and data complexity. To evaluate the model’s performance, a comprehensive comparative analysis is conducted, involving five algorithms: Artificial Neural Network (ANN), BiLSTM, GRU, hybrid BiLSTM-GRU, and Transformer models. The findings indicate that employing the TIR model leads to superior accuracy in the analyzed area, achieving R2 value of 0.9983, RMSE of 0.0140, and MAE of 0.0092. This performance surpasses those of the alternative models studied. The integration of BiLSTM and GRU algorithms with the attention mechanism and positional encoding has been optimized to enhance the forecasting of SI. This approach mitigates computational dependencies and minimizes the error terms within the model.

Sparse-Coding Variational Autoencoders.

The sparse coding model posits that the visual system has evolved to efficiently code natural stimuli using a sparse set of features from an overcomplete dictionary. The original sparse coding model suffered from two key limitations, however: (1) computing the neural response to an image patch required minimizing a nonlinear objective function via recurrent dynamics and (2) fitting relied on approximate inference methods that ignored uncertainty. Although subsequent work has developed several methods to overcome these obstacles, we propose a novel solution inspired by the variational autoencoder (VAE) framework. We introduce the sparse coding variational autoencoder (SVAE), which augments the sparse coding model with a probabilistic recognition model parameterized by a deep neural network. This recognition model provides a neurally plausible feedforward implementation for the mapping from image patches to neural activities and enables a principled method for fitting the sparse coding model to data via maximization of the evidence lower bound (ELBO). The SVAE differs from standard VAEs in three key respects: the latent representation is overcomplete (there are more latent dimensions than image pixels), the prior is sparse or heavy-tailed instead of gaussian, and the decoder network is a linear projection instead of a deep network. We fit the SVAE to natural image data under different assumed prior distributions and show that it obtains higher test performance than previous fitting methods. Finally, we examine the response properties of the recognition network and show that it captures important nonlinear properties of neurons in the early visual pathway.

Applicability domain of a calibration model based on neural networks and infrared spectroscopy

Artificial neural networks are used as calibration models in routine analytical determinations that involve spectroscopic data. To ensure that the model will generate reliable predictions for new samples, the applicability domain must be well defined. This article describes a strategy for establishing the limits of the applicability domain when the calibration model is a feed-forward neural network. The applicability domain was defined by two limits: 1) the 0.99 quantile of the squared Mahalanobis distance calculated from the network activations of the training set and 2) the 0.99 quantile of the reconstruction error of the training spectra using either an autoencoder network or a decoder network. A new sample with a squared Mahalanobis distance and/or spectral residuals beyond these limits is said to be outside the applicability domain, and the prediction is questionable. The approach was illustrated by predicting the density of diesel fuel samples from mid-infrared spectra and the fat content in meat from near-infrared spectra. The methodology could correctly detect anomalous spectra in prediction using either the autoencoder or the decoder.

Music-stylized hierarchical dance synthesis with user control

BackgroundSynthesizing dance motions to match musical inputs is a significant challenge in animation research. Compared to functional human motions, such as locomotion, dance motions are creative and artistic, often influenced by music, and can be independent body language expressions. Dance choreography requires motion content to follow a general dance genre, whereas dance performances under musical influence are infused with diverse impromptu motion styles. Considering the high expressiveness and variations in space and time, providing accessible and effective user control for tuning dance motion styles remains an open problem. MethodsIn this study, we present a hierarchical framework that decouples the dance synthesis task into independent modules. We use a high-level choreography module built as a Transformer-based sequence model to predict the long-term structure of a dance genre and a low-level realization module that implements dance stylization and synchronization to match the musical input or user preferences. This novel framework allows the individual modules to be trained separately. Because of the decoupling, dance composition can fully utilize existing high-quality dance datasets that do not have musical accompaniments, and the dance implementation can conveniently incorporate user controls and edit motions through a decoder network. Each module is replaceable at runtime, which adds flexibility to the synthesis of dance sequences. ResultsSynthesized results demonstrate that our framework generates high-quality diverse dance motions that are well adapted to varying musical conditions and user controls.

MLFA-UNet: A multi-level feature assembly UNet for medical image segmentation

Medical image segmentation is crucial for accurate diagnosis and treatment in medical image analysis. Among the various methods employed, fully convolutional networks (FCNs) have emerged as a prominent approach for segmenting medical images. Notably, the U-Net architecture and its variants have gained widespread adoption in this domain. This paper introduces MLFA-UNet, an innovative architectural framework aimed at advancing medical image segmentation. MLFA-UNet adopts a U-shaped architecture and integrates two pivotal modules: multi-level feature assembly (MLFA) and multi-scale information attention (MSIA), complemented by a pixel-vanishing (PV) attention mechanism. These modules synergistically contribute to the segmentation process enhancement, fostering both robustness and segmentation precision. MLFA operates within both the network encoder and decoder, facilitating the extraction of local information crucial for accurately segmenting lesions. Furthermore, the bottleneck MSIA module serves to replace stacking modules, thereby expanding the receptive field and augmenting feature diversity, fortified by the PV attention mechanism. These integrated mechanisms work together to boost segmentation performance by effectively capturing both detailed local features and a broader range of contextual information, enhancing both accuracy and resilience in identifying lesions. To assess the versatility of the network, we conducted evaluations of MFLA-UNet across a range of medical image segmentation datasets, encompassing diverse imaging modalities such as wireless capsule endoscopy (WCE), colonoscopy, and dermoscopic images. Our results consistently demonstrate that MFLA-UNet outperforms state-of-the-art algorithms, achieving dice coefficients of 91.42%, 82.43%, 90.8%, and 88.68% for the MICCAI 2017 (Red Lesion), ISIC 2017, PH2, and CVC-ClinicalDB datasets, respectively.