Comparison of Network Architectures for Semantic Segmentation

Research Square (Research Square), Journal Year: 2024, Volume and Issue: unknown

Published: June 14, 2024

Semantic segmentation is an important task in computer vision applications like autonomous driving, remote sensing, and medical image processing. Recently deep convolutional neural networks have become a standard semantic tasks. The encoder-decoder architecture, example of which the UNet, has well established widely used. However, two stream architecture working at different resolutions, DDRNet, shown promising results. This paper aims to compare performance two-stream by training UNet DDRNet under same conditions over datasets. results this showed that all cases, with without pre-training, datasets outperformed architecture. Although established, seems inferior when compared newer DDRNet.

Language: Английский

MPCCN: A Symmetry-Based Multi-Scale Position-Aware Cyclic Convolutional Network for Retinal Vessel Segmentation

Symmetry, Journal Year: 2024, Volume and Issue: 16(9), P. 1189 - 1189

Published: Sept. 10, 2024

In medical image analysis, precise retinal vessel segmentation is crucial for diagnosing and managing ocular diseases as the vascular network reflects numerous health indicators. Despite decades of development, challenges such intricate textures, ruptures, undetected areas persist, particularly in accurately segmenting small vessels addressing low contrast imaging. This study introduces a novel approach called MPCCN that combines position-aware cyclic convolution (PCC) with multi-scale resolution input to tackle these challenges. By integrating standard PCC, effectively captures both global local features. A module enhances feature extraction, while weighted-shared residual guided attention minimizes background noise emphasizes structures. Our achieves sensitivity values 98.87%, 99.17%, 98.88%; specificity 98.93%, 97.25%, 99.20%; accuracy scores 97.38%, 97.85%, 97.75%; AUC 98.90%, 99.15%, 99.05% on DRIVE, STARE, CHASE_DB1 datasets, respectively. addition, it records F1 90.93%, 91.00%, 90.55%. Experimental results demonstrate our method outperforms existing techniques, especially detecting vessels.

Language: Английский

A multi-scale feature extraction and fusion-based model for retinal vessel segmentation in fundus images

Medical & Biological Engineering & Computing, Journal Year: 2024, Volume and Issue: unknown

Published: Oct. 21, 2024

Language: Английский