Multi-scale conv-attention U-Net for medical image segmentation DOI Creative Commons
Linqiang Pan, Chengxue Zhang, Jingbo Sun

и другие.

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Апрель 8, 2025

U-Net-based network structures are widely used in medical image segmentation. However, effectively capturing multi-scale features and spatial context information of complex organizational remains a challenge. To address this, we propose novel structure based on the U-Net backbone. This model integrates Adaptive Convolution (AC) module, Multi-Scale Learning (MSL) Conv-Attention module to enhance feature expression ability segmentation performance. The AC dynamically adjusts convolutional kernel through an adaptive layer. enables extract different shapes scales adaptively, further improving its performance scenarios. MSL is designed for fusion. It aggregates fine-grained high-level semantic from resolutions, creating rich connections between encoding decoding processes. On other hand, incorporates efficient attention mechanism into skip connections. captures global using low-dimensional proxy high-dimensional data. approach reduces computational complexity while maintaining effective channel extraction. Experimental validation CVC-ClinicDB, MICCAI 2023 Tooth, ISIC2017 datasets demonstrates that our proposed MSCA-UNet significantly improves accuracy robustness. At same time, it lightweight outperforms existing methods.

Язык: Английский

Liver segmentation network based on detail enhancement and multi-scale feature fusion DOI Creative Commons

Lu Tinglan,

Jun Qin, Guihe Qin

и другие.

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Янв. 3, 2025

Due to the low contrast of abdominal CT (Computer Tomography) images and similar color shape liver other organs such as spleen, stomach, kidneys, segmentation presents significant challenges. Additionally, 2D obtained from different angles (such sagittal, coronal, transverse planes) increase diversity morphology complexity segmentation. To address these issues, this paper proposes a Detail Enhanced Convolution (DE Conv) improve feature learning thereby enhance performance. Furthermore, enable model better learn features at scales, Multi-Scale Feature Fusion module (MSFF) is added skip connections in model. The MSFF enhances capture global features, thus improving accuracy Through aforementioned research, network based on detail enhancement multi-scale fusion (DEMF-Net). We conducted extensive experiments LiTS17 dataset, results demonstrate that DEMF-Net achieved improvements across various evaluation metrics. Therefore, proposed can achieve precise

Язык: Английский

Процитировано

2
Encoder-Free Multiaxis Physics-Aware Fusion Network for Remote Sensing Image Dehazing DOI
Yuanbo Wen, Tao Gao, Jing Zhang

и другие.

IEEE Transactions on Geoscience and Remote Sensing, Год журнала: 2023, Номер 61, С. 1 - 15

Опубликована: Янв. 1, 2023

Current methods for remote sensing image dehazing confront noteworthy computational intricacies and yield suboptimal dehazed outputs, thereby circumscribing their pragmatic applicability. To this end, we propose EMPF-Net, a novel encoder-free multi-axis physics-aware fusion network that exhibits both light-weighted characteristics efficiency. In our pipeline, contend conventional u-shaped networks allocate substantial resources to encode haze-degraded features, which play subordinate role in the reconstruction process. Consequently, encoder stages solely incorporate down-sampling operations. improve representation efficiency enhance generalization capabilities, devise partial queried learning block (MPQLB) primarily concentrates on dimension-wise queries, instead of relying strictly-correlated content input features. Furthermore, augment procedure by incorporating ground truth supervision into each stage via supervised cross-scale transposed attention module (SCTAM). It calculates maps under guidance clean images, suppressing less informative features propagate subsequent level. addition, address challenge ineffective intral-level feature fusion, result insufficient elimination information negatively impact quality reconstructed introduce intra-level (PIFM). This harnesses physical inversion model facilitate interaction alleviate interference dehazing-irrelevant information. Our proposed EMPF-Net is evaluated 12 publicly available datasets, experimental results substantiate superiority terms metrical scores visual quality, despite being equipped with modest parameter count 300 K. approach readily accessible at https://github.com/chdwyb/EMPF-Net.

Язык: Английский

Процитировано

28
Rolling-Unet: Revitalizing MLP’s Ability to Efficiently Extract Long-Distance Dependencies for Medical Image Segmentation DOI Open Access
Yutong Liu,

Haijiang Zhu,

Mengting Liu

и другие.

Proceedings of the AAAI Conference on Artificial Intelligence, Год журнала: 2024, Номер 38(4), С. 3819 - 3827

Опубликована: Март 24, 2024

Medical image segmentation methods based on deep learning network are mainly divided into CNN and Transformer. However, struggles to capture long-distance dependencies, while Transformer suffers from high computational complexity poor local feature learning. To efficiently extract fuse features long-range this paper proposes Rolling-Unet, which is a model combined with MLP. Specifically, we propose the core R-MLP module, responsible for dependency in single direction of whole image. By controlling combining modules different directions, OR-MLP DOR-MLP formed dependencies multiple directions. Further, Lo2 block proposed encode both context information without excessive burden. has same parameter size as 3×3 convolution. The experimental results four public datasets show that Rolling-Unet achieves superior performance compared state-of-the-art methods.

Язык: Английский

Процитировано

9
PACAF-Net: pixel shuffling based fiderality-preserved up/downsampling and adaptive cross-attention fusion for effective medical image segmentation DOI

Yuanhang Cai,

Aouaidjia Kamel, Chongsheng Zhang

и другие.

Signal Image and Video Processing, Год журнала: 2025, Номер 19(5)

Опубликована: Март 3, 2025

Язык: Английский

Процитировано

1
PAMSNet: A medical image segmentation network based on spatial pyramid and attention mechanism DOI
Yuncong Feng, Xiaoyan Zhu, Xiaoli Zhang

и другие.

Biomedical Signal Processing and Control, Год журнала: 2024, Номер 94, С. 106285 - 106285

Опубликована: Апрель 1, 2024

Язык: Английский

Процитировано

8
UCM-Net: A lightweight and efficient solution for skin lesion segmentation using MLP and CNN DOI
Chunyu Yuan, Dongfang Zhao, Sos С. Agaian

и другие.

Biomedical Signal Processing and Control, Год журнала: 2024, Номер 96, С. 106573 - 106573

Опубликована: Июль 3, 2024

Язык: Английский

Процитировано

8
DSEUNet: A lightweight UNet for dynamic space grouping enhancement for skin lesion segmentation DOI
Jian Li, Jiawei Wang,

Fengwu Lin

и другие.

Expert Systems with Applications, Год журнала: 2024, Номер 255, С. 124544 - 124544

Опубликована: Дек. 1, 2024

Язык: Английский

Процитировано

6
MugenNet: A Novel Combined Convolution Neural Network and Transformer Network with Application in Colonic Polyp Image Segmentation DOI Creative Commons
Chen Peng, Zhiqin Qian,

Kunyu Wang

и другие.

Sensors, Год журнала: 2024, Номер 24(23), С. 7473 - 7473

Опубликована: Ноя. 23, 2024

Accurate polyp image segmentation is of great significance, because it can help in the detection polyps. Convolutional neural network (CNN) a common automatic method, but its main disadvantage long training time. Transformer another method that be adapted to by employing self-attention mechanism, which essentially assigns different importance weights each piece information, thus achieving high computational efficiency during segmentation. However, potential drawback with risk information loss. The study reported this paper employed well-known hybridization principle propose combine CNN and retain strengths both. Specifically, applied early colonic polyps implement model called MugenNet for We conducted comprehensive experiment compare other models on five publicly available datasets. An ablation was as well. experimental results showed achieve mean Dice 0.714 ETIS dataset, optimal performance dataset compared models, an inference speed 56 FPS. overall outcome optimally two methods machine learning are complementary other.

Язык: Английский

Процитировано

5
MSDEnet: Multi-scale detail enhanced network based on human visual system for medical image segmentation DOI Creative Commons
Yuangang Ma, Hong Xu, Yue Feng

и другие.

Computers in Biology and Medicine, Год журнала: 2024, Номер 170, С. 108010 - 108010

Опубликована: Янв. 18, 2024

In medical image segmentation, accuracy is commonly high for tasks involving clear boundary partitioning features, as seen in the segmentation of X-ray images. However, objects with less obvious such skin regions similar color textures or CT images adjacent organs Hounsfield value ranges, significantly decreases. Inspired by human visual system, we proposed multi-scale detail enhanced network. Firstly, designed a module to enhance contrast between central and peripheral receptive field information using superposition two asymmetric convolutions different directions standard convolution. Then, expanded scale into module. The difference at scales makes network more sensitive changes details, resulting accurate segmentation. order reduce impact redundant on results increase effective field, channel module, adapted from Res2net This creates independent parallel branches within single residual structure, increasing utilization sensory level. We conducted experiments four datasets, our method outperformed common algorithms currently being used. Additionally, carried out detailed ablation confirm effectiveness each

Язык: Английский

Процитировано

4
SCSONet: spatial-channel synergistic optimization net for skin lesion segmentation DOI Creative Commons
Haoyu Chen, Zexin Li,

Xinyue Huang

и другие.

Frontiers in Physics, Год журнала: 2024, Номер 12

Опубликована: Март 20, 2024

In the field of computer-assisted medical diagnosis, developing image segmentation models that are both accurate and capable real-time operation under limited computational resources is crucial. Particularly for skin disease segmentation, construction such lightweight must balance cost efficiency, especially in environments with computing power, memory, storage. This study proposes a new network designed specifically aimed at significantly reducing number parameters floating-point operations while ensuring performance. The proposed ConvStem module, full-dimensional attention, learns complementary attention weights across all four dimensions convolution kernel, effectively enhancing recognition irregularly shaped lesion areas, model’s parameter count burden, thus promoting model lightweighting performance improvement. SCF Block reduces feature redundancy through spatial channel fusion, lowering improving results. paper validates effectiveness robustness SCSONet on two public datasets, demonstrating its low resource requirements. https://github.com/Haoyu1Chen/SCSONet .

Язык: Английский

Процитировано

4