计算机科学
人工智能
分割
图像(数学)
相关性
模式识别(心理学)
计算机视觉
图像分割
图像融合
融合
统计模型
尺度空间分割
数据挖掘
传感器融合
相关系数
统计分析
图像处理
人工神经网络
作者
Shuanghua Ye,Wenwen Tang,Huang Zhao,Huiyu Zhou,Zhiwen Zheng,Jin Liu,Xiaoshuai Zhang,Xingru Huang
标识
DOI:10.1016/j.eswa.2026.133860
摘要
Accurate volumetric segmentation relies on encoder-decoder skip connections to recover spatial details. Standard U-shaped networks fuse shallow encoder features with deep decoder features by direct concatenation. This assumes statistical compatibility between streams. In multi-scale and multi-modal volumes this assumption often fails. Feature moments drift across modalities and scales. The mismatch propagates modality-dependent noise and downsampling artifacts into the decoder and weakens boundary refinement. We propose Statistical Correlation Reconciliation Network (SCoReN) which treats each skip junction as a calibration interface. Its core Duplex inter-Scale Statistical Concordance operator (DiSCo) performs bidirectional reconciliation. DiSCo injects compressed high-resolution evidence into the low-resolution stream through cross-scale token concordance. It reinforces axis-aware continuity in the high-resolution stream with gated axial enhancement. It then applies correlation conditioned moment gates and mixture gates to suppress statistically discordant channels before fusion. We evaluate on ten public datasets spanning CT, MRI and OCT and benchmark against 20 recent state-of-the-art (SoTA) 3D segmentation methods covering convolutional backbones transformer variants and state-space models. SCoReN improves overlap and surface accuracy across heterogeneous targets. The model uses 13.66M parameters and 166.63G FLOPs. These results suggest that statistically calibrated skip fusion may provide an efficient route for robust volumetric segmentation under heterogeneous imaging conditions. Our code is publicly available at https://github.com/IMOP-lab/SCoReN .
科研通智能强力驱动
Strongly Powered by AbleSci AI