人工智能
分割
计算机科学
深静脉
斑点图案
特征(语言学)
深度学习
模式识别(心理学)
一般化
计算机视觉
散斑噪声
可转让性
人工神经网络
卷积神经网络
血栓形成
图像分割
深层神经网络
融合
放射科
病变
掷骰子
边界(拓扑)
超声波
水准点(测量)
图像(数学)
Sørensen–骰子系数
作者
Meiyun Chen,Qilong Tan,Haixia He,Lishan Huang,Wen Xi,Dawei Zhang
出处
期刊:Physica Scripta
[IOP Publishing]
日期:2026-06-12
卷期号:101 (23): 236003-236003
标识
DOI:10.1088/1402-4896/ae7774
摘要
Deep vein thrombosis (DVT) poses a life-threatening risk due to potential pulmonary embolism, making timely and accurate diagnosis critical. However, automated segmentation of DVT in ultrasound images remains highly challenging due to multi-scale lesion features, pervasive speckle noise, and indistinct boundaries. To address these challenges, we propose dual-stream cross-branch fusion network (DSCF-Net), a dual-stream CNN-Mamba network with cross-branch fusion specifically designed for DVT segmentation. In the encoder, the CNN branch extracts fine-grained local features while the Mamba branch captures long-range dependencies to handle multi-scale lesion variations. The dual-stream spatial attention fusion (DSAF) module performs adaptive refinement to mitigate boundary ambiguity, while the cross pooling-guided fusion (CPGF) module employs multi-scale feature aggregation to suppress speckle noise. Experimental results on the clinical DVT dataset demonstrate that DSCF-Net achieves a Dice score of 78.27%, MIoU of 65.15%, F1 of 78.17%, and HD95 of 15.23 mm, outperforming state-of-the-art methods. Generalization experiments on the public TN3K dataset also achieved state-of-the-art performance, confirming the method’s strong transferability in ultrasound image segmentation.
科研通智能强力驱动
Strongly Powered by AbleSci AI