Synthesizing Images With Annotations for Medical Image Segmentation Using Diffusion Probabilistic Model DOI Creative Commons

Zengan Huang,

Qinzhu Yang, Mu Tian

et al.

International Journal of Imaging Systems and Technology, Journal Year: 2024, Volume and Issue: 35(1)

Published: Dec. 14, 2024

ABSTRACT To alleviate the burden of manual annotation, there are numerous excellent segmentation models for images being developed. However, performance these data‐driven is frequently constrained by availability samples sizes pair medical and annotations. Therefore, to address this challenge, study introduces image augmentation diffusion model (MEDSAD). MEDSAD solves problem annotation scarcity utilizing a given simple generate paired images. improve stability, we used traditional study. exert better control over texture synthesis in generated MEDSAD, style injection (TSI) mechanism introduced. Additionally, propose feature frequency domain attention (FFDA) module mitigate adverse effects high‐frequency noise during generation. The efficacy substantiated through validation three distinct tasks encompassing magnetic resonance (MR) ultrasound (US) imaging modalities, focusing on breast tumors, brain nerve structures. findings demonstrate model's proficiency synthesizing pairs based provided annotations, thereby facilitating notable subsequent tasks. Moreover, improvement becomes greater as quantity synthetic available data increases. This underscores robust generalization capability intrinsic model, potentially offering avenues future explorations training research.

Language: Английский

Retinal multi-lesion segmentation by reinforcing single-lesion guidance with multi-view learning DOI
Liyun Zhang, Zhiwen Fang, Ting Li

et al.

Biomedical Signal Processing and Control, Journal Year: 2023, Volume and Issue: 86, P. 105349 - 105349

Published: Aug. 12, 2023

Language: Английский

Citations

4
Architectures and Applications of U-net in Medical Image Segmentation: A Review DOI
Jundi Wang, Lei Han, Dongsheng Ran

et al.

Published: June 1, 2023

Recently, with the increasing application of deep learning in medical field, convolutional neural networks, represented by U-Net, has been widely applied image segmentation. The improved U-shaped network structure based on U-Net gradually become a hot topic segmentation research. This article summarizes improvement work related to from three perspectives: modifying skip connections, adding or replacing blocks and concatenating multiple networks. Then, taking retina, lungs, brain, abdomen, other organs as examples, characteristics difficulties various organ were introduced. Finally, summary outlook made.

Language: Английский

Citations

4
Self-supervised pre-training for joint optic disc and cup segmentation via attention-aware network DOI Creative Commons
Zhi‐Wang Zhou,

Yuanchang Zheng,

Xiaoyu Zhou

et al.

BMC Ophthalmology, Journal Year: 2024, Volume and Issue: 24(1)

Published: March 4, 2024

Abstract Image segmentation is a fundamental task in deep learning, which able to analyse the essence of images for further development. However, supervised learning method, collecting pixel-level labels very time-consuming and labour-intensive. In medical image processing area optic disc cup segmentation, we consider there are two challenging problems that remain unsolved. One how design an efficient network capture global field execute fast real applications. The other train using few training data due some privacy issues. this paper, conquer such issues, first novel attention-aware model equipped with multi-scale attention module pyramid structure-like encoder-decoder network, can efficiently learn semantics long-range dependencies input images. Furthermore, also inject prior knowledge lies inside by loss function. Then, propose self-supervised contrastive method segmentation. unsupervised feature representation learned matching encoded query dictionary keys technique. Finetuning pre-trained proposed function help achieve good performance task. To validate effectiveness extensive systemic evaluations on different public benchmarks, including DRISHTI-GS REFUGE datasets demonstrate superiority new state-of-the-art approaching 0.9801 0.9087 F 1 score respectively while gaining 0.9657 $$DC_{disc}$$ D C disc 0.8976 $$DC_{cup}$$ cup . code will be made publicly available.

Language: Английский

Citations

1
Deep learning technology in vascular image segmentation and disease diagnosis DOI Creative Commons

Chengyang Du,

Jie Zhuang, Xinglu Huang

et al.

Journal of intelligent medicine., Journal Year: 2024, Volume and Issue: unknown

Published: Sept. 24, 2024

Abstract Blood vessel segmentation is a crucial aspect of medical image processing, aiding professionals in more accurate disease analysis and diagnosis. Manual blood methods are time‐consuming cumbersome, making the development automatic essential. The rapid advancements deep learning technology have introduced new tools for vascular segmentation. In this review, we provide comprehensive overview learning‐based across various fields, including retinal segmentation, cerebrovascular pulmonary Several prevalent diseases, such as tumors, posed significant health challenges globally. This review also discusses application diagnosis within these contexts. Finally, considering current research landscape, discuss existing potential future developments We aim to assist researchers gaining understanding designing effective models, ultimately offering opportunities early treatment.

Language: Английский

Citations

1
Synthesizing Images With Annotations for Medical Image Segmentation Using Diffusion Probabilistic Model DOI Creative Commons

Zengan Huang,

Qinzhu Yang, Mu Tian

et al.

International Journal of Imaging Systems and Technology, Journal Year: 2024, Volume and Issue: 35(1)

Published: Dec. 14, 2024

ABSTRACT To alleviate the burden of manual annotation, there are numerous excellent segmentation models for images being developed. However, performance these data‐driven is frequently constrained by availability samples sizes pair medical and annotations. Therefore, to address this challenge, study introduces image augmentation diffusion model (MEDSAD). MEDSAD solves problem annotation scarcity utilizing a given simple generate paired images. improve stability, we used traditional study. exert better control over texture synthesis in generated MEDSAD, style injection (TSI) mechanism introduced. Additionally, propose feature frequency domain attention (FFDA) module mitigate adverse effects high‐frequency noise during generation. The efficacy substantiated through validation three distinct tasks encompassing magnetic resonance (MR) ultrasound (US) imaging modalities, focusing on breast tumors, brain nerve structures. findings demonstrate model's proficiency synthesizing pairs based provided annotations, thereby facilitating notable subsequent tasks. Moreover, improvement becomes greater as quantity synthetic available data increases. This underscores robust generalization capability intrinsic model, potentially offering avenues future explorations training research.

Language: Английский

Citations

1