TP-Net: Two-Path Network for Retinal Vessel Segmentation

IEEE Journal of Biomedical and Health Informatics, Journal Year: 2023, Volume and Issue: 27(4), P. 1979 - 1990

Published: Jan. 17, 2023

Refined and automatic retinal vessel segmentation is crucial for computer-aided early diagnosis of retinopathy. However, existing methods often suffer from mis-segmentation when dealing with thin low-contrast vessels. In this paper, a two-path network proposed, namely TP-Net, which consists three core parts, i.e. main-path, sub-path, multi-scale feature aggregation module (MFAM). Main-path to detect the trunk area vessels, sub-path effectively capture edge information The prediction results two paths are combined by MFAM, obtaining refined three-layer lightweight backbone elaborately designed according characteristics then global selection mechanism (GFSM) can autonomously select features that more important task at different layers network, thereby, enhancing capability an extraction method loss function enhance ability reduce Finally, MFAM proposed fuse main-path remove background noises while preserving details, thus, TP-Net has been evaluated on public datasets, DRIVE, STARE, CHASE DB1. experimental show achieved superior performance generalization fewer model parameters compared state-of-the-art methods.

Language: Английский

---

Stimulus-guided adaptive transformer network for retinal blood vessel segmentation in fundus images

Medical Image Analysis, Journal Year: 2023, Volume and Issue: 89, P. 102929 - 102929

Published: Aug. 9, 2023

Automated retinal blood vessel segmentation in fundus images provides important evidence to ophthalmologists coping with prevalent ocular diseases an efficient and non-invasive way. However, segmenting vessels is a challenging task, due the high variety scale appearance of similarity visual features between lesions vascular. Inspired by way that cortex adaptively responds type stimulus, we propose Stimulus-Guided Adaptive Transformer Network (SGAT-Net) for accurate segmentation. It entails Module (SGA-Module) can extract local-global compound based on inductive bias self-attention mechanism. Alongside light-weight residual encoder (ResEncoder) structure capturing relevant details appearance, Pooling (SGAP-Former) introduced reweight maximum average pooling enrich contextual embedding representation while suppressing redundant information. Moreover, Feature Fusion (SGAFF) module designed emphasize local global context fuse them latent space adjust receptive field (RF) task. The evaluation implemented largest image dataset (FIVES) three popular datasets (DRIVE, STARE, CHASEDB1). Experimental results show proposed method achieves competitive performance over other existing method, clear advantage avoiding errors commonly happen areas highly similar features. sourcecode publicly available at: https://github.com/Gins-07/SGAT.

Language: Английский

---

SGU-Net: Shape-Guided Ultralight Network for Abdominal Image Segmentation

IEEE Journal of Biomedical and Health Informatics, Journal Year: 2023, Volume and Issue: 27(3), P. 1431 - 1442

Published: Jan. 19, 2023

Convolutional neural networks (CNNs) have achieved significant success in medical image segmentation. However, they also suffer from the requirement of a large number parameters, leading to difficulty deploying CNNs low-source hardwares, e.g., embedded systems and mobile devices. Although some compacted or small memory-hungry models been reported, most them may cause degradation segmentation accuracy. To address this issue, we propose shape-guided ultralight network (SGU-Net) with extremely low computational costs. The proposed SGU-Net includes two main contributions: it first presents an convolution that is able implement double separable convolutions simultaneously, i.e., asymmetric depthwise convolution. not only effectively reduces parameters but enhances robustness SGU-Net. Secondly, our SGUNet employs additional adversarial shape-constraint let learn shape representation targets, which can significantly improve accuracy for abdomen images using self-supervision. extensively tested on four public benchmark datasets, LiTS, CHAOS, NIH-TCIA 3Dircbdb. Experimental results show achieves higher lower memory costs, outperforms state-of-the-art networks. Moreover, apply into 3D volume network, obtains comparable performance fewer usage. available code released at https://github.com/SUST-reynole/SGUNet.

Language: Английский

---

TCC-net: A two-stage training method with contradictory loss and co-teaching based on meta-learning for learning with noisy labels

Information Sciences, Journal Year: 2023, Volume and Issue: 639, P. 119008 - 119008

Published: April 28, 2023

Language: Английский

---

StrokeNet: An automated approach for segmentation and rupture risk prediction of intracranial aneurysm

Computerized Medical Imaging and Graphics, Journal Year: 2023, Volume and Issue: 108, P. 102271 - 102271

Published: July 22, 2023

Language: Английский

---

Deep learning for retinal vessel segmentation: a systematic review of techniques and applications

Medical & Biological Engineering & Computing, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 18, 2025

Language: Английский

---

GDSSA-Net: A gradually deeply supervised self-ensemble attention network for IoMT-integrated thyroid nodule segmentation

Internet of Things, Journal Year: 2025, Volume and Issue: unknown, P. 101598 - 101598

Published: April 1, 2025

Language: Английский

---

A novel sea-land segmentation network for enhanced coastline extraction using satellite remote sensing images

Advances in Space Research, Journal Year: 2024, Volume and Issue: 74(5), P. 2200 - 2213

Published: June 14, 2024

Language: Английский

---

H-EMD: A Hierarchical Earth Mover’s Distance Method for Instance Segmentation

IEEE Transactions on Medical Imaging, Journal Year: 2022, Volume and Issue: 41(10), P. 2582 - 2597

Published: April 21, 2022

Deep learning (DL) based semantic segmentation methods have achieved excellent performance in biomedical image segmentation, producing high quality probability maps to allow extraction of rich instance information facilitate good segmentation. While numerous efforts were put into developing new DL models, less attention was paid a key issue how effectively explore their attain the best possible We observe that by models can be used generate many candidates, and accurate selecting from them set "optimized" candidates as output instances. Further, generated form well-behaved hierarchical structure (a forest), which allows instances an optimized manner. Hence, we propose novel framework, called earth mover's distance (H-EMD), for 2D+time videos 3D images, judiciously incorporates consistent selection with semantic-segmentation-generated maps. H-EMD contains two main stages: (1) candidate generation: capturing instance-structured generating forest structure; (2) selection: final formulate problem on optimization (EMD), solve it integer linear programming. Extensive experiments eight video or datasets demonstrate consistently boosts is highly competitive state-of-the-art methods.

Language: Английский

---

DI-UNet: dual-branch interactive U-Net for skin cancer image segmentation

Journal of Cancer Research and Clinical Oncology, Journal Year: 2023, Volume and Issue: 149(17), P. 15511 - 15524

Published: Aug. 30, 2023

Language: Английский

---