Core Concepts
提案されたMDU-Netは、多様なスケールの特徴マップを組み合わせることで、生体医用画像セグメンテーションの性能を向上させます。
Abstract
Abstract:
MDU-Net proposes multi-scale dense connections for biomedical image segmentation.
Fuses feature maps from higher and lower layers to enhance feature propagation.
Quantization improves segmentation performance.
Introduction:
Biomedical image segmentation importance in medical diagnosis and treatment.
Traditional methods time-consuming, automatic segmentation needed.
Related Work:
U-Net architecture widely used for semantic segmentation tasks.
Dense connections improve feature reuse and fusion in networks.
Method:
Three multi-scale dense connections introduced for encoder, decoder, and across them.
Detailed experiments on different dense connection structures.
Experiments:
Evaluation on GlaS dataset shows improved accuracy with increasing number of dense connections.
Efficiency comparison with other networks shows MDU-Net's high efficiency.
Conclusion:
MDU-Net combines three dense connected architectures to enhance feature representation.
Achieves superior dice coefficient over U-net by up to 4.1% on test B.
Stats
MDU-Netは、MICCAI 2015 Gland Segmentation(GlaS)データセットにおいて、テストAで最大1.8%、テストBで最大3.5%の改善を達成しました。
Quotes
"Biomedical image segmentation plays a central role in quantitative analysis, clinical diagnosis, and medical intervention."
"Accurate automatic medical image segmentation attracts people’s attention and has wide application prospects."