Multi-scale Feature Fusion Module Research Articles

MTC-Net: Multi-scale feature fusion network for medical image segmentation

Image segmentation is critical in medical image processing for lesion detection, localisation, and subsequent diagnosis. Currently, computer-aided diagnosis (CAD) has played a significant role in improving diagnostic efficiency and accuracy. The segmentation task is made more difficult by the hazy lesion boundaries and uneven forms. Because standard convolutional neural networks (CNNs) are incapable of capturing global contextual information, adequate segmentation results are impossible to achieve. We propose a multiscale feature fusion network (MTC-Net) in this paper that integrates deep separable convolution and self-attentive modules in the encoder to achieve better local continuity of images and feature maps. In the decoder, a multi-branch multi-scale feature fusion module (MSFB) is utilized to improve the network’s feature extraction capability, and it is integrated with a global cooperative aggregation module (GCAM) to learn more contextual information and adaptively fuse multi-scale features. To develop rich hierarchical representations of irregular forms, the suggested detail enhancement module (DEM) adaptively integrates local characteristics with their global dependencies. To validate the effectiveness of the proposed network, we conducted extensive experiments, evaluated on the public datasets of skin, breast, thyroid and gastrointestinal tract with ISIC2018, BUSI, TN3K and Kvasir-SEG. The comparison with the latest methods also verifies the superiority of our proposed MTC-Net in terms of accuracy. Our code on https://github.com/gih23/MTC-Net.

MCDNet: Multilevel cloud detection network for remote sensing images based on dual-perspective change-guided and multi-scale feature fusion

Cloud detection plays a crucial role in the preprocessing of optical remote sensing images. While extensive deep learning-based methods have shown strong performance in detecting thick clouds, their ability to identify thin and broken clouds is often inadequate due to their sparse distribution, semi-transparency, and similarity to background regions. To address this limitation, we introduce a multilevel cloud detection network (MCDNet) capable of simultaneously detecting thick and thin clouds. This network effectively enhances the accuracy of identifying thin and broken clouds by integrating a dual-perspective change-guided mechanism (DPCG) and a multi-scale feature fusion module (MSFF). The DPCG creates a dual-input stream by combining the original image with the thin cloud removal image, and then utilizes a dual-perspective feature fusion module (DPFF) to perform feature fusion and extract change features, thereby improving the model's ability to perceive thin cloud regions and mitigate inter-class similarity in multilevel cloud detection. The MSFF enhances the model's sensitivity to broken clouds by utilizing multiple non-adjacent low-level features to remedy the missing spatial information in the high-level features during multiple downsampling. Experimental results on the L8-Biome and WHUS2-CD datasets demonstrate that MCDNet significantly enhances the detection performance of both thin and broken clouds, and outperforms state-of-the-art methods in accuracy and efficiency. The code of MCDNet is available in https://github.com/djw-easy/MCDNet.

MAD-UNet: A Multi-Region UAV Remote Sensing Network for Rural Building Extraction.

For the development of an idyllic rural landscape, an accurate survey of rural buildings is essential. The extraction of rural structures from unmanned aerial vehicle (UAV) remote sensing imagery is prone to errors such as misclassifications, omissions, and subpar edge detailing. This study introduces a multi-scale fusion and detail enhancement network for rural building extraction, termed the Multi-Attention-Detail U-shaped Network (MAD-UNet). Initially, an atrous convolutional pyramid pooling module is integrated between the encoder and decoder to enhance the main network's ability to identify buildings of varying sizes, thereby reducing omissions. Additionally, a Multi-scale Feature Fusion Module (MFFM) is constructed within the decoder, utilizing superficial detail features to refine the layered detail information, which improves the extraction of small-sized structures and their edges. A coordination attention mechanism and deep supervision modules are simultaneously incorporated to minimize misclassifications. MAD-UNet has been tested on a private UAV building dataset and the publicly available Wuhan University (WHU) Building Dataset and benchmarked against models such as U-Net, PSPNet, DeepLabV3+, HRNet, ISANet, and AGSCNet, achieving Intersection over Union (IoU) scores of 77.43% and 91.02%, respectively. The results demonstrate its effectiveness in extracting rural buildings from UAV remote sensing images across different regions.

Self-supervised multi-magnification feature enhancement for segmentation of hepatocellular carcinoma region in pathological images

Accurate segmentation of hepatocellular carcinoma regions is a critical step in surgical evaluation of whole slide pathological images. Recently, emerging multi-magnification learning-based methods have shown promise with evaluation of whole slide images. However, traditional multi-magnification learning segmentation models focus on uniformity and fail to effectively exploit different magnification information. The purpose of this study is to develop a novel multi-magnification learning segmentation model by effectively utilizing diverse magnification information. Therefore, we developed a novel multi-magnification feature self-enhancement network to extract magnification-specific features from multi-magnification images via self-supervised learning without extra supervision. Specifically, the proposed network enhances feature information via a super-resolution module based on self-supervision to transfer the representative information into a segmentation encoder, which takes full advantage of the pyramid storage characteristic of whole slide images. Moreover, a multi-scale feature fusion module based on the attention mechanism is used to fuse pretrained multi-scale features through a gated unit block, which is designed to fuse features from generation task into and segmentation task. We evaluated this method using the hepatocellular carcinoma data set from The Cancer Genome Atlas with cross-validation and achieved a 0.829 Dice similarity coefficient, nearly a 3% improvement over state-of-the-art models used for segmentation. To the best of our knowledge, this is the first study of the application of super-resolution as a self-enhancement network for histopathological image segmentation via self-supervised pretext learning. The code will be available at https://github.com/SH-Diao123/Self-Supervised-Multi-magnification-Segmentation.

SoybeanTracer: An In-Field scene property-based framework for high-throughput soybean canopy coverage extraction and evaluation

Soybean is a crucial plant-based protein and vegetable oil source for the global population and a significant cereal-oil dual-purpose crop. Extracting soybean canopy coverage, subsequently implementing evaluation in a low-cost, high-throughput and accurate manner, is highly valuable for soybean breeding and yield increase. Although the utilization of RGB remote sensing images for soybean analysis offers distinct advantages in terms of its cost effectiveness, there are still issues to be addressed. Specifically, the in-field scene properties of soybean canopy data have rarely been developed and there is a gap in automatic evaluation of soybean canopy coverage. For purpose of tackling these pending issues, this study proposed SoybeanTracer, a novel framework designed to achieve accurate soybean canopy coverage extraction and enable specialized evaluation with low cost and high throughput. The proposed framework comprised two components: Soybean Canopy Segmentation Network (SCSNet) and an agronomic soybean coverage evaluation system. Specifically, SCSNet employed a feature extraction network integrating the homogeneity within plot regions and soybean distribution characteristics. To further address the segmentation challenge of soybean plants with sparse growth posed by genotypes, a soybean multiscale feature fusion module was designed. As a result, SCSNet achieved a mean intersection over union (mIoU) and mean pixel accuracy (mPA) of 89.39% and 94.42%, respectively, on soybean canopy segmentation. Therefore, accurate and low-cost extraction of soybean canopy coverage was achieved through the use of RGB remote sensing data. Furthermore, the soybean canopy coverage evaluation system proposed in this study encompassed five levels, characterized by wider spacing at the two ends and narrower spacing in the middle. This agronomic evaluation system had been effectively implemented on 196 plots. The variance in the pairwise levels of the coverage evaluation results demonstrated the robustness of the evaluation system. SoybeanTracer could indeed help to extract soybean canopy coverage and provide valuable reference for implementing soybean genotype selection, laying a useful foundation for other crops in similar scientific tasks. The source code and trained models were available at https://github.com/KLoAS-Research/SCSNet-SoybeanTracer.

Image Classification Method Based on Multi-Scale Convolutional Neural Network

Traditional convolutional neural networks (CNNs) typically use fixed scale convolutional kernels for feature extraction when processing image classification tasks, while ignoring the multi-scale information present in the image. To overcome this limitation, we propose an algorithm based on multi-scale CNNs, which capture features at different levels by introducing convolutional kernels of different scales into the convolutional layer. In this study, we first designed a multi-scale convolutional layer consisting of multiple convolutional kernels of different scales to extract multi-scale features of the image. To further enhance classification performance, we introduced a multi-scale feature fusion module that can effectively fuse features of different scales and classify them through a fully connected layer. Then we conducted extensive experiments on several commonly used image classification datasets. The experimental results show that this network can not only effectively identify and locate hyperspectral image targets in different scenarios, but also reduce missed detections and false positives during the detection process. The average accuracy of the improved model has been improved, and the recognition accuracy of some small markers affected by external factors such as occlusion and lighting has also been improved. In addition, by comparing the detection effect of a single image, the progressiveness and anti-leakage ability of the improved model are proved. The image classification method based on multi-scale CNNs has broad application prospects in image recognition and feature extraction, and can provide valuable reference and reference for research in related fields.

MUSCLE: multi-view and multi-scale attentional feature fusion for microRNA-disease associations prediction.

MicroRNAs (miRNAs) synergize with various biomolecules in human cells resulting in diverse functions in regulating a wide range of biological processes. Predicting potential disease-associated miRNAs as valuable biomarkers contributes to the treatment of human diseases. However, few previous methods take a holistic perspective and only concentrate on isolated miRNA and disease objects, thereby ignoring that human cells are responsible for multiple relationships. In this work, we first constructed a multi-view graph based on the relationships between miRNAs and various biomolecules, and then utilized graph attention neural network to learn the graph topology features of miRNAs and diseases for each view. Next, we added an attention mechanism again, and developed a multi-scale feature fusion module, aiming to determine the optimal fusion results for the multi-view topology features of miRNAs and diseases. In addition, the prior attribute knowledge of miRNAs and diseases was simultaneously added to achieve better prediction results and solve the cold start problem. Finally, the learned miRNA and disease representations were then concatenated and fed into a multi-layer perceptron for end-to-end training and predicting potential miRNA-disease associations. To assess the efficacy of our model (called MUSCLE), we performed 5- and 10-fold cross-validation (CV), which got average the Area under ROC curves of 0.966${\pm }$0.0102 and 0.973${\pm }$0.0135, respectively, outperforming most current state-of-the-art models. We then examined the impact of crucial parameters on prediction performance and performed ablation experiments on the feature combination and model architecture. Furthermore, the case studies about colon cancer, lung cancer and breast cancer also fully demonstrate the good inductive capability of MUSCLE. Our data and code are free available at a public GitHub repository: https://github.com/zht-code/MUSCLE.git.

MEF-UNet: An end-to-end ultrasound image segmentation algorithm based on multi-scale feature extraction and fusion

Ultrasound image segmentation is a challenging task due to the complexity of lesion types, fuzzy boundaries, and low-contrast images along with the presence of noises and artifacts. To address these issues, we propose an end-to-end multi-scale feature extraction and fusion network (MEF-UNet) for the automatic segmentation of ultrasound images. Specifically, we first design a selective feature extraction encoder, including detail extraction stage and structure extraction stage, to precisely capture the edge details and overall shape features of the lesions. In order to enhance the representation capacity of contextual information, we develop a context information storage module in the skip-connection section, responsible for integrating information from adjacent two-layer feature maps. In addition, we design a multi-scale feature fusion module in the decoder section to merge feature maps with different scales. Experimental results indicate that our MEF-UNet can significantly improve the segmentation results in both quantitative analysis and visual effects.

MRFA-Net: Kidney Segmentation Method Based on Multi-Scale Feature Fusion and Residual Full Attention

For the characterization of the kidney segmentation task, this paper proposes a self-supervised kidney segmentation method based on multi-scale feature fusion and residual full attention, named MRFA-Net. In this study, we introduce the multi-scale feature fusion module to extract multi-scale information of kidneys from abdominal CT slices; additionally, the residual full-attention convolution module is designed to handle the multi-scale information of kidneys by introducing a full-attention mechanism, thus improving the segmentation results of kidneys. The Dice coefficient on the Kits19 dataset reaches 0.972. The experimental results demonstrate that the proposed method achieves good segmentation performance compared to other algorithms, effectively enhancing the accuracy of kidney segmentation.

A feature enhancement FCOS algorithm for dynamic traffic object detection

The development of object detection plays an important role in the realisation of fully autonomous driving, and the feature extraction is the key step for object detection. There has been significant difference and scale variation of object features for different road traffic participants (RTPs), meanwhile traditional Convolutional Neural Networks (CNNs) was difficult to extract object features efficiently for small targets. In order to improve the ability of feature extraction, a RTP object detection method combining dynamic convolution and feature enhancement was proposed. The Fully Convolutional One-Stage (FCOS) object detection algorithm was used as baseline. First, the dynamic convolution module was designed in the backbone network to identify different object features to the maximum extent. Second, a dual attention module was designed to filter object feature information while reducing the amount of computation. Finally, in the detection part, the feature expression ability of shallow network was further enhanced by multi-scale feature fusion module, and the effectiveness of the proposed algorithm was verified using Cityscapes dataset. The experimental result indicated that mAP increased by 2.3% compared with baseline. This study can improve the efficiency of RTP detection and contribute to the industrialisation of intelligent connected vehicles.

A novel difficult-to-segment samples focusing network for oral CBCT image segmentation

Using deep learning technology to segment oral CBCT images for clinical diagnosis and treatment is one of the important research directions in the field of clinical dentistry. However, the blurred contour and the scale difference limit the segmentation accuracy of the crown edge and the root part of the current methods, making these regions become difficult-to-segment samples in the oral CBCT segmentation task. Aiming at the above problems, this work proposed a Difficult-to-Segment Focus Network (DSFNet) for segmenting oral CBCT images. The network utilizes a Feature Capturing Module (FCM) to efficiently capture local and long-range features, enhancing the feature extraction performance. Additionally, a Multi-Scale Feature Fusion Module (MFFM) is employed to merge multiscale feature information. To further improve the loss ratio for difficult-to-segment samples, a hybrid loss function is proposed, combining Focal Loss and Dice Loss. By utilizing the hybrid loss function, DSFNet achieves 91.85% Dice Similarity Coefficient (DSC) and 0.216 mm Average Symmetric Surface Distance (ASSD) performance in oral CBCT segmentation tasks. Experimental results show that the proposed method is superior to current dental CBCT image segmentation techniques and has real-world applicability.

Hyperspectral image change detection based on an improved multi-scale and spectral-wise transformer

ABSTRACT Change detection (CD) has a significant application in the remote sensing field. Because of the popularity of hyperspectral image (HSI) and the application of deep learning methods, hyperspectral image change detection (HSI-CD) techniques have been greatly developed. Among them, convolutional neural network (CNN) has garnered the greatest interest in HSI-CD due to their superior feature learning capabilities. However, current CNN-based algorithms have trouble capturing spectral similarity and long-range dependency owing to their intrinsic structural restrictions. Recently, transformers have been shown to extract global dependency from nature images in an extremely efficient way. But it has some difficulties in handling high-dimensional data, such as HSI. To address these issues, we propose an improved multi-scale and spectral-wise transformer (MS-SWT). The proposed MS-SWT is capable of capturing spectral similarity and long-range dependence between bands to enhance the efficiency of the HSI-CD task. Furthermore, to maximize the utilization of spatial information, we present a multi-scale feature fusion module (MFFM) to extract and fuse different dimensions of spatial features. More importantly, a locality self-attention (LSA) module is employed to alleviate the problem of smoothing the distribution of attention scores due to the large number of spectral embeddings. Moreover, we design a channel self-supervised loss function that can capture intrinsic information from the spectral channels to further strengthen the robustness of model training when the training samples are scarce. Lastly, comprehensive experiments present the high performance of our MS-SWT on four bitemporal HSI datasets and demonstrate the superiority of MS-SWT over state-of-the-art approaches.

Multi-scale feature learning via residual dynamic graph convolutional network for hyperspectral image classification

ABSTRACT Thanks to the relevant spectral-spatial information, hyperspectral images (HSIs) have been widely exploited in Earth observation. Recently, graph convolutional networks (GCNs) have attracted increasing attention in HSI classification due to their advantages in processing non-Euclidean structure data. Unlike convolutional neural networks (CNNs), which perform convolution operations on regular square regions, GCNs can directly work on graph structure data to extract the relationships among adjacent land covers. However, extracting meaningful and deep discriminative spectral-spatial features from HSIs is still a challenging task. In this article, a novel multi-scale feature learning via residual dynamic graph convolutional network is designed for HSI classification, which can extract large-scale contextual spatial structures at superpixel-level graph and local spectral-spatial information at pixel-level, significantly improving the performance of HSI classification. Unlike from the existing GCN-based methods that operate on a graph with a fixed neighbourhood size, multiple graphs with diverse neighbourhood scales are built to comprehensively leverage spectral-spatial information and relationship at multiple scales, and these graphs are dynamically updated to generate more discriminative features (via dynamic GCN) during the convolution process. Moreover, to fully use the multi-scale features extracted from HSIs, a multi-scale feature fusion module is developed to emphasize important features and suppress irrelevant ones. Extensive experiments carried on three benchmark data sets demonstrate the superiority of the proposed approach over other state-of-the-art methods.

Swin-Net: A Swin-Transformer-Based Network Combing with Multi-Scale Features for Segmentation of Breast Tumor Ultrasound Images

Breast cancer is one of the most common cancers in the world, especially among women. Breast tumor segmentation is a key step in the identification and localization of the breast tumor region, which has important clinical significance. Inspired by the swin-transformer model with powerful global modeling ability, we propose a semantic segmentation framework named Swin-Net for breast ultrasound images, which combines Transformer and Convolutional Neural Networks (CNNs) to effectively improve the accuracy of breast ultrasound segmentation. Firstly, our model utilizes a swin-transformer encoder with stronger learning ability, which can extract features of images more precisely. In addition, two new modules are introduced in our method, including the feature refinement and enhancement module (RLM) and the hierarchical multi-scale feature fusion module (HFM), given that the influence of ultrasonic image acquisition methods and the characteristics of tumor lesions is difficult to capture. Among them, the RLM module is used to further refine and enhance the feature map learned by the transformer encoder. The HFM module is used to process multi-scale high-level semantic features and low-level details, so as to achieve effective cross-layer feature fusion, suppress noise, and improve model segmentation performance. Experimental results show that Swin-Net performs significantly better than the most advanced methods on the two public benchmark datasets. In particular, it achieves an absolute improvement of 1.4–1.8% on Dice. Additionally, we provide a new dataset of breast ultrasound images on which we test the effect of our model, further demonstrating the validity of our method. In summary, the proposed Swin-Net framework makes significant advancements in breast ultrasound image segmentation, providing valuable exploration for research and applications in this domain.

HmsU-Net: A hybrid multi-scale U-net based on a CNN and transformer for medical image segmentation

Accurate medical image segmentation is of great significance for subsequent diagnosis and analysis. The acquisition of multi-scale information plays an important role in segmenting regions of interest of different sizes. With the emergence of Transformers, numerous networks adopted hybrid structures incorporating Transformers and CNNs to learn multi-scale information. However, the majority of research has focused on the design and composition of CNN and Transformer structures, neglecting the inconsistencies in feature learning between Transformer and CNN. This oversight has resulted in the hybrid network's performance not being fully realized. In this work, we proposed a novel hybrid multi-scale segmentation network named HmsU-Net, which effectively fused multi-scale features. Specifically, HmsU-Net employed a parallel design incorporating both CNN and Transformer architectures. To address the inconsistency in feature learning between CNN and Transformer within the same stage, we proposed the multi-scale feature fusion module. For feature fusion across different stages, we introduced the cross-attention module. Comprehensive experiments conducted on various datasets demonstrate that our approach surpasses current state-of-the-art methods.

A Multiscale Delayed Channel Attention Network-Based Method for Pansharpening

Nowadays, remote sensing images are widely used in many fields. To obtain high-quality remote sensing images, remote sensing image fusion methods have attracted much attention. Although convolutional neural network-based pansharpening methods have good results, these methods focus on the forward propagation of the network, which cannot effectively seek the mapping relationship between images. Moreover, it is difficult to obtain global information due to the limitations of convolutional operations. In this paper, we propose a pansharpening method based on multiscale delayed channel attention networks. The method iteratively seeks to correlate high-resolution stage features with the original low-resolution multispectral image, providing a mechanism for error feedback to map the error of each stage. Meanwhile, it designs a multiscale feature fusion module to fuse feature information from different fields of view. The design of the delayed channel attention mechanism makes the network acquire the association relationship between low-frequency information and high-frequency information through adaptive learning, giving different weights to high-frequency information, and making it more flexible in dealing with different types of information. Finally, the feature aggregation module is used to generate fused images and adjust the corresponding feature information. The experimental results obtained from the Gaofen-2 and WorldView-2 experimental data show that the method achieves a significant improvement compared to the current fusion algorithms.

DFNet-Trans: An end-to-end multibranching network for depth estimation for transparent objects

Transparent objects play a vital role in modern industries and find widespread applications across various engineering scenarios. However, capturing accurate depth maps of transparent objects remains challenging due to their reflective and refractive properties, which pose difficulties for most commercial-grade optical sensors. In this paper, we propose a novel depth estimation method called DFNet-Trans, designed to estimate depth from a noisy RGB-D image input. Initially, a multiscale feature fusion module (FFM) is incorporated into the existing depth estimation network to generate the initial depth map. Subsequently, we enhance the network by adding a confidence branch and a mask branch on the same encoder, enabling improved distortion correction and real scene restoration in the depth estimation. Based on the framework representation, missing depth can be completed. Comprehensive experiments demonstrate that the proposed approach significantly outperforms the current state-of-the-art methods on the recently popular large-scale real dataset TransCG. the proposed approach achieves a remarkable 27.7% reduction in RMSE and a notable 34.6% reduction in REL. The generalization experiment shows that the proposed approach outperforms existing methods when generalized to an unknown real dataset.

Synthetic aperture optical image restoration based on multi-scale feature enhancement

With the wide applications of high-resolution imaging technology in topographic mapping, astronomical observation, and military reconnaissance and other fields, the requirements for imaging resolution of optical system are becoming higher and higher . According to the diffraction limit and Rayleigh criterion, the imaging resolution of the optical system is proportional to the size of the aperture of the system, but affected by the material and the processing of the optical component: the single aperture of the optical system cannot be infinitely enlarged. Therefore the synthetic aperture technology is proposed to replace the single large aperture optical system. Owing to the effect of sub-aperture arrangement and light scattering, the imaging of synthetic aperture optical system will be degraded because of insufficient light area and phase distortion. The traditional imaging restoration algorithm of synthetic aperture optical system is sensitive to noise, overly relies on degraded model, requires a lot of manually designed models, and has poor adaptability. To solve this problem, a multi-scale feature enhancement method of restoring the synthetic aperture optical image is proposed in this work. U-Net is used to obtain multi-scale feature, and self-attention in mixed domain is used to improve the ability of of the network to extract the features in space and channel. Multi-scale feature fusion module and feature enhancement module are constructed to fuse the information between features on different scales. The information interaction mode of the codec layer is optimized, the attention of the whole network to the real structure of the original image is enhanced, and the artifact interference caused by ringing is avoided in the process of restoration. The final experimental results are 1.51%, 4.42% and 5.22% higher than those from the advanced deep learning algorithms in the evaluation indexes of peak signal-to-noise ratio, structural similarity and perceived similarity, respectively. In addition, the method presented in this work has a good restoration effect on the degraded images to different degrees of synthetic aperture, and can effectively restore the degraded images and the images with abnormal light, so as to solve the problem of imaging degradation of synthetic aperture optical system. The feasibility of deep learning method in synthetic aperture optical image restoration is proved.

Integration and Innovation of Higher Education Management and Student Training Mechanisms Based on Multi-scale Feature Fusion

Abstract In the current educational reform wave, which prioritizes moral development as the fundamental goal and stresses the enhancement of core student competencies, aligning higher education management with student cultivation mechanisms is crucial to educational progress. The Lucas-Kanade (LK) optical flow algorithm is used in this paper to gather behavioral characteristics from student interactions within the cultivation mechanism. A multi-scale convolutional kernel approach is used to fuse these features both locally and globally. A multi-scale feature fusion module subsequently classifies and recognizes these features, with recognition accuracy optimized by a tailored loss function. This approach allows for the timely identification and analysis of students’ aberrant behaviors, which aids in prompt educational interventions. The multi-scale feature fusion model can effectively identify various types of aberrant student behaviors, which aids educators and institutional leaders in their management efforts, as revealed by experimental findings. Statistically, the model’s implementation led to significant improvements in classroom routine compliance, with pre-and post-test p-values in the experimental cohort showing a notable difference (p = 0.001; p < 0.05). The proposed multi-scale feature fusion model promotes the integration of management and training mechanisms in higher education while also supporting the development of students’ learning capabilities. This innovation sets a solid foundation for future educational advancements.

MAG-Net : Multi-fusion network with grouped attention for retinal vessel segmentation.

Retinal vessel segmentation plays a vital role in the clinical diagnosis of ophthalmic diseases. Despite convolutional neural networks (CNNs) excelling in this task, challenges persist, such as restricted receptive fields and information loss from downsampling. To address these issues, we propose a new multi-fusion network with grouped attention (MAG-Net). First, we introduce a hybrid convolutional fusion module instead of the original encoding block to learn more feature information by expanding the receptive field. Additionally, the grouped attention enhancement module uses high-level features to guide low-level features and facilitates detailed information transmission through skip connections. Finally, the multi-scale feature fusion module aggregates features at different scales, effectively reducing information loss during decoder upsampling. To evaluate the performance of the MAG-Net, we conducted experiments on three widely used retinal datasets: DRIVE, CHASE and STARE. The results demonstrate remarkable segmentation accuracy, specificity and Dice coefficients. Specifically, the MAG-Net achieved segmentation accuracy values of 0.9708, 0.9773 and 0.9743, specificity values of 0.9836, 0.9875 and 0.9906 and Dice coefficients of 0.8576, 0.8069 and 0.8228, respectively. The experimental results demonstrate that our method outperforms existing segmentation methods exhibiting superior performance and segmentation outcomes.