Scholarlink

<p>Despite notable progress in existing deep learning algorithms, fetal abdominal circumference segmentation in ultrasound images remains challenging due to low image quality and complex acoustics. In this paper, we present a novel and effective segmentation framework, termed CGFI-Net, to address the aforementioned challenges. Specifically, MobileNetV2 is adopted as the encoder backbone to efficiently extract hierarchical features while maintaining low computational complexity. To enhance the representation capability of skip connections, a spatial-attention guided multi-scale dilated fusion module (SA-MSDFM) is introduced to capture contextual information across multiple receptive fields and emphasize anatomically relevant regions. In the decoding stage, we design a cross-gated feature interaction module (CGFIM) to adaptively fuse skip-connection features from three different encoder levels, enabling effective cross-level feature interaction and information refinement. Furthermore, a cascaded residual feature aggregation module (CRFAM) is employed at the output stage to progressively integrate multi-level features and enhance boundary delineation. Finally, CGFI-Net was evaluated on a fetal abdominal circumference ultrasound dataset collected from Quzhou People’s Hospital, where it achieved Accuracy of 0.9929, Dice of 0.9736, Recall of 0.9728, Mcc of 0.9695, and Jaccard of 0.9486. To further validate the generalization capability of CGFI-Net, additional experiments were conducted on the publicly available digital database thyroid image (DDTI), yielding Accuracy, Dice, Recall, Mcc, and Jaccard scores of 0.9465, 0.8095, 0.7926, 0.7803, and 0.6820. Compared with the selected representative methods, CGFI-Net achieves superior segmentation performance while maintaining high computational efficiency, requiring only 3.21 GFLOPs and 6.10M parameters, and achieving an inference speed exceeding 70 FPS on an NVIDIA GeForce RTX 4090 GPU, demonstrating its potential adaptability for practical clinical applications.</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

Cross-gated feature interaction network for fetal abdominal circumference image segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月1日

查看全文
  • Yan Cheng,
  • Xiaokang Ding,
  • Jinfeng You

摘要

Despite notable progress in existing deep learning algorithms, fetal abdominal circumference segmentation in ultrasound images remains challenging due to low image quality and complex acoustics. In this paper, we present a novel and effective segmentation framework, termed CGFI-Net, to address the aforementioned challenges. Specifically, MobileNetV2 is adopted as the encoder backbone to efficiently extract hierarchical features while maintaining low computational complexity. To enhance the representation capability of skip connections, a spatial-attention guided multi-scale dilated fusion module (SA-MSDFM) is introduced to capture contextual information across multiple receptive fields and emphasize anatomically relevant regions. In the decoding stage, we design a cross-gated feature interaction module (CGFIM) to adaptively fuse skip-connection features from three different encoder levels, enabling effective cross-level feature interaction and information refinement. Furthermore, a cascaded residual feature aggregation module (CRFAM) is employed at the output stage to progressively integrate multi-level features and enhance boundary delineation. Finally, CGFI-Net was evaluated on a fetal abdominal circumference ultrasound dataset collected from Quzhou People’s Hospital, where it achieved Accuracy of 0.9929, Dice of 0.9736, Recall of 0.9728, Mcc of 0.9695, and Jaccard of 0.9486. To further validate the generalization capability of CGFI-Net, additional experiments were conducted on the publicly available digital database thyroid image (DDTI), yielding Accuracy, Dice, Recall, Mcc, and Jaccard scores of 0.9465, 0.8095, 0.7926, 0.7803, and 0.6820. Compared with the selected representative methods, CGFI-Net achieves superior segmentation performance while maintaining high computational efficiency, requiring only 3.21 GFLOPs and 6.10M parameters, and achieving an inference speed exceeding 70 FPS on an NVIDIA GeForce RTX 4090 GPU, demonstrating its potential adaptability for practical clinical applications.