MANet: a multi-attention network for automatic liver tumor segmentation in computed tomography (CT) imaging

Scientific Reports, Год журнала: 2023, Номер 13(1)

Опубликована: Ноя. 16, 2023

Automatic liver tumor segmentation is a paramount important application for diagnosis and treatment planning. However, it has become highly challenging task due to the heterogeneity of shape intensity variation. capable establish diagnostic standard provide relevant radiological information all levels expertise. Recently, deep convolutional neural networks have demonstrated superiority in feature extraction learning medical image segmentation. multi-layer dense stacks make model quite inconsistent imitating visual attention awareness expertise recognition task. To bridge that capability, mechanisms developed better selection. In this paper, we propose novel network named Multi Attention Network (MANet) as fusion learn highlighting features while suppressing irrelevant The proposed followed U-Net basic architecture. Moreover, residual mechanism implemented encoder. Convolutional block module split into channel spatial modules implement encoder decoder integrated extract low-level combine with high-level ones. architecture trained evaluated on publicly available MICCAI 2017 Liver Tumor Segmentation dataset 3DIRCADb under various evaluation metrics. MANet promising results compared state-of-the-art methods comparatively small parameter overhead.

Язык: Английский

---

A Semantic Segmentation Method Based on AS-Unet++ for Power Remote Sensing of Images

Sensors, Год журнала: 2024, Номер 24(1), С. 269 - 269

Опубликована: Янв. 2, 2024

In order to achieve the automatic planning of power transmission lines, a key step is precisely recognize feature information remote sensing images. Considering that has different depths and distribution not uniform, semantic segmentation method based on new AS-Unet++ proposed in this paper. First, atrous spatial pyramid pooling (ASPP) squeeze-and-excitation (SE) module are added traditional Unet, such field can be expanded important features enhanced, which called AS-Unet. Second, an structure built by using layers AS-Unet, extraction parts each layer AS-Unet stacked together. Compared with automatically learns at determines depth optimal performance. Once number network determined, excess pruned, will greatly reduce trained parameters. The experimental results show overall recognition accuracy significantly improved compared Unet.

Язык: Английский

---

Reading recognition of pointer meters based on an improved UNet++ network

Measurement Science and Technology, Год журнала: 2023, Номер 35(3), С. 035009 - 035009

Опубликована: Дек. 4, 2023

Abstract Pointer meters are widely used in modern industries, such as petrochemical applications, substations, and nuclear power plants. To overcome the reading errors inaccurate measurements due to uneven or fluctuating illumination practical this paper proposes an improved UNet++ network for recognizing pointer meter readings. First, scale invariant feature transform feature-matching algorithm is adjust captured tilted images a symmetrical upright shape. Then, segment regions dashboard eliminate background interference. Furthermore, part of convolution replaced with dilated different expansion rates expand perceptual field during training. In jump connection, attention mechanism module also introduced path enhance region’s features be segmented suppress parts non-segmented area. A hybrid loss function model training prevent imbalance region share. Finally, distance method read gauge representation. Experiments were conducted compare performance proposed that original terms feasibility precision. The experimental results showed recognition accuracy was significantly by enhanced network, accuracy, sensitivity, specificity reaching 98.65%, 84.33%, 99.38%, respectively. when using numerical reading, average relative error only 0.122%, indicating its robustness natural environment.

Язык: Английский

---

Swin-Unet++: a study on phenotypic parameter analysis of cabbage seedling roots

Plant Methods, Год журнала: 2025, Номер 21(1)

Опубликована: Март 3, 2025

As an important economic crop, the growth status of root system cabbage directly affects its overall health and yield. To monitor seedlings during their period, this study proposes a new network architecture called Swin-Unet++. This integrates Swin-Transformer module residual networks uses attention mechanisms to replace traditional convolution operations for feature extraction. It also adopts concept fuse contextual information from different levels, addressing issue insufficient extraction thin mesh-like roots seedlings. Compared with other backbone high-precision semantic segmentation networks, SwinUnet + achieves superior results. The results show that accuracy Swin-Unet in tasks reached as high 98.19%, model parameter 60 M average response time 29.5 ms. classic Unet network, mIoU increased by 1.08%, verifying can accurately extract fine-grained features roots. Furthermore, when images after segmentations are compared locate position through contours, has best positioning effect. On basis pixels obtained segmentation, calculated maximum length, extension width, thickness actual measurements. resulting goodness fit R² values 94.82%, 94.43%, 86.45%, respectively. Verifying effectiveness extracting phenotypic traits seedling framework developed provides technique monitoring analysis systems, ultimately leading development automated platform offers technical support intelligent agriculture efficient planting practices.

Язык: Английский

---

Res-ECA-UNet++: an automatic segmentation model for ovarian tumor ultrasound images based on residual networks and channel attention mechanism

Frontiers in Medicine, Год журнала: 2025, Номер 12

Опубликована: Май 21, 2025

Ultrasound imaging has emerged as the preferred modality for ovarian tumor screening due to its non-invasive nature and real-time dynamic capabilities. However, in many developing countries, ultrasound diagnosis remains dependent on specialist physicians, where shortage of skilled professionals relatively low accuracy manual diagnoses significantly constrain efficiency. Although deep learning achieved remarkable progress medical image segmentation recent years, existing methods still face challenges segmentation, including insufficient robustness, imprecise boundary delineation, dependence high-performance hardware facilities. This study proposes a learning-based automatic model, Res-ECA-UNet++, designed enhance while alleviating strain limited healthcare resources. The Res-ECA-UNet++ model employs UNet++ fundamental architecture with ResNet34 serving backbone network. To effectively address vanishing gradient problem networks, residual modules are incorporated into skip connections between encoding decoding processes. integration enhances feature extraction efficiency improving stability generalization Furthermore, ECA-Net channel attention mechanism is introduced during downsampling phase. adaptively emphasizes region-related information through global recalibration, thereby recognition localization precision areas. Based clinical datasets tumors, experimental results demonstrate that achieves outstanding performance validation, Dice coefficient 95.63%, mean Intersection over Union (mIoU) 91.84%, 99.75%. Compared baseline UNet, improves these three metrics by 0.45, 4.42, 1.57%, respectively. Comparative analyses ROC curves AUC values further indicate exhibits superior enhanced capabilities test set. In terms computational efficiency, inference time meets requirements both high-end low-end hardware, demonstrating suitability deployment resource-constrained devices. Additionally, comparative experiments public OTU2D dataset validate model's performance, highlighting strong potential practical applications. proposed demonstrates exceptional robustness images, application. Its ability aid clinicians underscores broad prospects implementation. Future research will focus optimizing improve adaptability diverse pathological types characteristics, expanding diagnostic utility.

Язык: Английский

---

CotepRes-Net: An efficient U-Net based deep learning method of liver segmentation from Computed Tomography images

Biomedical Signal Processing and Control, Год журнала: 2023, Номер 88, С. 105660 - 105660

Опубликована: Ноя. 3, 2023

Язык: Английский

---

Mask Guidance Pyramid Network for Overlapping Cervical Cell Edge Detection

Applied Sciences, Год журнала: 2023, Номер 13(13), С. 7526 - 7526

Опубликована: Июнь 26, 2023

An important indicator of cervical cancer diagnosis is to calculate the proportion diseased cells and cells, so it necessary segment judge cell status. The existing methods are difficult deal with segmentation overlapping cells. In order solve this problem, we put forward such a hypothesis by reading large number literature, that is, image edge measurement tasks have unity in high-level features. To prove hypothesis, paper, focus on complementary between information object get higher accuracy detection results. Specifically, present joint multi-task learning framework for mask guidance pyramid network. main component Mask Guidance Module (MGM), which integrates two stores shared latent semantics interact tasks. For semantic detection, propose novel Refinement Aggregated (RAM) fusion promote edges. Finally, improve pixel accuracy, consistency constraint loss function introduced our model training. Our extensive experiments proved method outperforms other efforts.

Язык: Английский

---

Dil-UNet++: A Multi-Scale Fusion Retinal Vessel Segmentation Network Model Based on UNet++

Computer Science and Application, Год журнала: 2024, Номер 14(01), С. 54 - 67

Опубликована: Янв. 1, 2024

Язык: Английский

---

Semantic segmentation method for micro-cracks in silicon nitride ceramic bearing balls based on coupling of edge channel enhancement and weighted gated attention mechanism in EMU-Net+

Measurement, Год журнала: 2024, Номер 238, С. 115333 - 115333

Опубликована: Июль 20, 2024

Язык: Английский

---

Deep learning assisted quantitative detection of cardiac troponin I in hierarchical dendritic copper–nickel nanostructure lateral flow immunoassay

Analytical Methods, Год журнала: 2024, Номер 16(39), С. 6715 - 6725

Опубликована: Янв. 1, 2024

This paper proposes a deep learning-based method using an improved UNet++ network with attention and residual modules to enhance quantitative detection accuracy in HD-nanoMetal LFIA sensor images.

Язык: Английский

---