PFFNet: A pyramid feature fusion network for microaneurysm segmentation in fundus images DOI Creative Commons

Jiaxin Lu,

Beiji Zou,

Xiaoxia Xiao

et al.

IET Image Processing, Journal Year: 2024, Volume and Issue: unknown

Published: Nov. 27, 2024

Abstract Retinal microaneurysm (MA) is a definite earliest clinical sigh of diabetic retinopathy (DR). Its automatic segmentation key to realizing intelligent screening for early DR, which can significantly reduce the risk visual impairment in patients. However, minute scale and subtle contrast MAs against background pose challenges segmentation. This paper focuses on MA fundus images. A novel pyramid feature fusion network (PFFNet) that progressively develops fuses rich contextual information by integrating two modules proposed. Multiple global scene parsing (GPSP) are introduced between encoder decoder provide diverse through reconstructing skip connections. Additionally, spatial scale‐aware (SSAP) module dynamically fuse multi‐scale information. will help identify from low‐contrast background. Furthermore, mitigate issue related category imbalance, combo loss function introduced. Finally, validate effectiveness proposed method, experiments conducted publicly available datasets, IDRiD DDR, PFFNet compared with several state‐of‐the‐art models. The experimental results demonstrate superiority our task.

Language: Английский

An ensemble approach of deep CNN models with Beta normalization aggregation for gastrointestinal disease detection DOI

Zafran Waheed,

Jinsong Gui,

Kamran Amjad

et al.

Biomedical Signal Processing and Control, Journal Year: 2025, Volume and Issue: 105, P. 107567 - 107567

Published: Feb. 4, 2025

Language: Английский

Citations

1

Alternate encoder and dual decoder CNN-Transformer networks for medical image segmentation DOI Creative Commons
Lin Zhang, Xinyu Guo,

Hongkun Sun

et al.

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: March 14, 2025

Accurately extracting lesions from medical images is a fundamental but challenging problem in image analysis. In recent years, methods based on convolutional neural networks and Transformer have achieved great success the segmentation field. Combining powerful perception of local information by CNNs efficient capture global context crucial for segmentation. However, unique characteristics many lesion tissues often lead to poor performance most previous models failed fully extract effective features. Therefore, an encoder-decoder architecture, we propose novel alternate encoder dual decoder CNN-Transformer network, AD2Former, with two attractive designs: 1) We alternating learning can achieve real-time interaction between information, allowing both mutually guide learning. 2) architecture. The way dual-branch independent decoding fusion. To efficiently fuse different feature sub-decoders during decoding, introduce channel attention module reduce redundant information. Driven these designs, AD2Former demonstrates strong ability target regions fuzzy boundaries. Experiments multi-organ skin datasets also demonstrate effectiveness superiority AD2Former.

Language: Английский

Citations

0

Monocular depth estimation via a detail semantic collaborative network for indoor scenes DOI Creative Commons
Wen Song, Xu Cui, Yakun Xie

et al.

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: March 31, 2025

Monocular image depth estimation is crucial for indoor scene reconstruction, and it plays a significant role in optimizing building energy efficiency, environment modeling, smart space design. However, the small variability of scenes leads to weakly distinguishable detail features. Meanwhile, there are diverse types objects, expression correlation among different objects complicated. Additionally, robustness recent models still needs further improvement given these environments. To address problems, detail‒semantic collaborative network (DSCNet) proposed monocular scenes. First, contextual features contained images fully captured via hierarchical transformer structure. Second, structure established, which establishes selective attention feature map extract details semantic information from maps. The extracted subsequently fused improve perception ability network. Finally, complex addressed by aggregating detailed at levels, model accuracy effectively improved without increasing number parameters. tested on NYU SUN datasets. approach produces state-of-the-art results compared with 14 performance optimal methods. In addition, discussed analyzed terms stability, robustness, ablation experiments availability

Language: Английский

Citations

0

FastUGI-Net: Enhanced real-time endoscopic diagnosis with efficient multi-task learning DOI
In Neng Chan, Pak Kin Wong, Tao Yan

et al.

Expert Systems with Applications, Journal Year: 2025, Volume and Issue: unknown, P. 127444 - 127444

Published: April 1, 2025

Language: Английский

Citations

0

A lighter hybrid feature fusion framework for polyp segmentation DOI Creative Commons
Xueqiu He,

Luo Yonggang,

Min Liu

et al.

Scientific Reports, Journal Year: 2024, Volume and Issue: 14(1)

Published: Oct. 5, 2024

Colonoscopy is widely recognized as the most effective method for detection of colon polyps, which crucial early screening colorectal cancer. Polyp identification and segmentation in colonoscopy images require specialized medical knowledge are often labor-intensive expensive. Deep learning provides an intelligent efficient approach polyp segmentation. However, variability size heterogeneity boundaries interiors pose challenges accurate Currently, Transformer-based methods have become a mainstream trend these tend to overlook local details due inherent characteristics Transformer, leading inferior results. Moreover, computational burden brought by self-attention mechanisms hinders practical application models. To address issues, we propose novel CNN-Transformer hybrid model (CTHP). CTHP combines strengths CNN, excels at modeling information, global semantics, enhance accuracy. We transform computation over entire feature map into width height directions, significantly improving efficiency. Additionally, design new information propagation module introduce additional positional bias coefficients during attention process, reduces dispersal introduced deep mixed fusion Transformer. Extensive experimental results demonstrate that our proposed achieves state-of-the-art performance on multiple benchmark datasets Furthermore, cross-domain generalization experiments show exhibits excellent performance.

Language: Английский

Citations

0

Automated lesion detection in gastrointestinal endoscopic images: leveraging deep belief networks and genetic algorithm-based Segmentation DOI
Mousa Alhajlah

Multimedia Tools and Applications, Journal Year: 2024, Volume and Issue: unknown

Published: Nov. 23, 2024

Language: Английский

Citations

0

PFFNet: A pyramid feature fusion network for microaneurysm segmentation in fundus images DOI Creative Commons

Jiaxin Lu,

Beiji Zou,

Xiaoxia Xiao

et al.

IET Image Processing, Journal Year: 2024, Volume and Issue: unknown

Published: Nov. 27, 2024

Abstract Retinal microaneurysm (MA) is a definite earliest clinical sigh of diabetic retinopathy (DR). Its automatic segmentation key to realizing intelligent screening for early DR, which can significantly reduce the risk visual impairment in patients. However, minute scale and subtle contrast MAs against background pose challenges segmentation. This paper focuses on MA fundus images. A novel pyramid feature fusion network (PFFNet) that progressively develops fuses rich contextual information by integrating two modules proposed. Multiple global scene parsing (GPSP) are introduced between encoder decoder provide diverse through reconstructing skip connections. Additionally, spatial scale‐aware (SSAP) module dynamically fuse multi‐scale information. will help identify from low‐contrast background. Furthermore, mitigate issue related category imbalance, combo loss function introduced. Finally, validate effectiveness proposed method, experiments conducted publicly available datasets, IDRiD DDR, PFFNet compared with several state‐of‐the‐art models. The experimental results demonstrate superiority our task.

Language: Английский

Citations

0