Scholarlink

<p>This study proposes a multi-scale dynamic enhancement solution to the clinical implementation bottleneck problem caused by data heterogeneity and small lesion characteristics in urinary system CT (Computed Tomography) diagnosis. A multi-center standardized CT urinary tract dataset is constructed at the data level, and double-blind annotation and elastic deformation enhancement are used. At the network architecture level, a dual-branch 3D convolutional network consisting of a ResNet-18 (Residual Network with 18 layers) plain scan branch and a DenseNet-121 (Densely Connected Convolutional Network with 121 layers) secretory branch is designed to achieve multi-phase feature fusion through a cross-attention mechanism. At the optimization level, an anatomical constraint attention module is introduced at the renal pelvis-ureter junction, and channel pruning is used to achieve lightweight deployment. Experimental results show that the lightweight dual-branch attention model in this paper maintains high performance (Dice 89.7%, IoU 82.9%) while reducing the computational cost to 48.8GFLOPs and the inference speed to 47.9ms. The model accuracy is improved by 1.6% compared with ViT-B/16, and the F1-score is improved by 3.2% compared with EfficientNet-B3. The lightweight dual-branch attention model in this paper can be applied to the clinical diagnosis process, improve the intelligent diagnosis level of urinary system CT images, and has a milestone significance for promoting the development of precision medicine.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Application of multi-scale dynamic enhancement based on deep neural network and CT urinary tract secretory phase image fusion in the diagnosis of urinary system diseases

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月4日

查看全文
  • Yuzhe Liang

摘要

This study proposes a multi-scale dynamic enhancement solution to the clinical implementation bottleneck problem caused by data heterogeneity and small lesion characteristics in urinary system CT (Computed Tomography) diagnosis. A multi-center standardized CT urinary tract dataset is constructed at the data level, and double-blind annotation and elastic deformation enhancement are used. At the network architecture level, a dual-branch 3D convolutional network consisting of a ResNet-18 (Residual Network with 18 layers) plain scan branch and a DenseNet-121 (Densely Connected Convolutional Network with 121 layers) secretory branch is designed to achieve multi-phase feature fusion through a cross-attention mechanism. At the optimization level, an anatomical constraint attention module is introduced at the renal pelvis-ureter junction, and channel pruning is used to achieve lightweight deployment. Experimental results show that the lightweight dual-branch attention model in this paper maintains high performance (Dice 89.7%, IoU 82.9%) while reducing the computational cost to 48.8GFLOPs and the inference speed to 47.9ms. The model accuracy is improved by 1.6% compared with ViT-B/16, and the F1-score is improved by 3.2% compared with EfficientNet-B3. The lightweight dual-branch attention model in this paper can be applied to the clinical diagnosis process, improve the intelligent diagnosis level of urinary system CT images, and has a milestone significance for promoting the development of precision medicine.