Scholarlink

Objectives <p>Placental segmentation and volumetry are crucial for clinical evaluation of maternal and fetal status. However, manual segmentation by physicians is time-consuming and demands specialized expertise. This study developed and evaluated a novel deep-learning algorithm to enhance placental segmentation and volumetry on magnetic resonance imaging (MRI).</p> Methods <p>This study utilized T2-weighted images from a single-center cohort of 101 pregnant women. The proposed placental segmentation model, CBGU-Net, was built upon the U-Net backbone. To improve performance, we integrated the convolutional block attention module (CBAM), boundary perception module (BPM), and attention gate (AG) into U-Net. The model's segmentation performance was quantitatively assessed using the dice similarity coefficient (DSC), pixel accuracy (ACC), and intersection over union (IoU). Clinical applicability was further evaluated by calculating the relative error in placental volumetry. Comparative experiments were performed against nnU-Net, U-Net, U-Net++, and DeepLabV3+ to benchmark the proposed model's performance.</p> Results <p>On the testing cohort, CBGU-Net achieved a mean DSC of 0.810 ± 0.181 (0.089 higher than U-Net++, <i>p</i> &lt; 0.05), a mean ACC of 0.986 ± 0.010 (0.004 higher than U-Net++, <i>p</i> &lt; 0.05), and a mean IoU of 0.694 ± 0.185 (0.099 higher than U-Net++, <i>p</i> &lt; 0.05). The relative volume measurement error of CBGU-Net on the testing cohort was 17.74 ± 12.33% (9.38% lower than that of nnU-Net,<i> p</i> &lt; 0.05).</p> Conclusion <p>The segmentation and volumetry results of the proposed CBGU-Net model on our dataset of T2-weighted images demonstrate promising placental segmentation performance and indicate its potential for clinical application.</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

Develop and evaluate deep learning models for automated MRI placental segmentation and volumetry

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Yechao Huang,
  • Joe Sarah Luciny,
  • Yumin Hu,
  • Dengfa Yang,
  • Pengjun Chen,
  • Zhanning Hu,
  • Zeliu Du,
  • Liyan Jiang,
  • Yongjin Zhou,
  • Shuiwei Xia,
  • Zufei Wang,
  • Chenying Lu,
  • Jiansong Ji

摘要

Objectives

Placental segmentation and volumetry are crucial for clinical evaluation of maternal and fetal status. However, manual segmentation by physicians is time-consuming and demands specialized expertise. This study developed and evaluated a novel deep-learning algorithm to enhance placental segmentation and volumetry on magnetic resonance imaging (MRI).

Methods

This study utilized T2-weighted images from a single-center cohort of 101 pregnant women. The proposed placental segmentation model, CBGU-Net, was built upon the U-Net backbone. To improve performance, we integrated the convolutional block attention module (CBAM), boundary perception module (BPM), and attention gate (AG) into U-Net. The model's segmentation performance was quantitatively assessed using the dice similarity coefficient (DSC), pixel accuracy (ACC), and intersection over union (IoU). Clinical applicability was further evaluated by calculating the relative error in placental volumetry. Comparative experiments were performed against nnU-Net, U-Net, U-Net++, and DeepLabV3+ to benchmark the proposed model's performance.

Results

On the testing cohort, CBGU-Net achieved a mean DSC of 0.810 ± 0.181 (0.089 higher than U-Net++, p < 0.05), a mean ACC of 0.986 ± 0.010 (0.004 higher than U-Net++, p < 0.05), and a mean IoU of 0.694 ± 0.185 (0.099 higher than U-Net++, p < 0.05). The relative volume measurement error of CBGU-Net on the testing cohort was 17.74 ± 12.33% (9.38% lower than that of nnU-Net, p < 0.05).

Conclusion

The segmentation and volumetry results of the proposed CBGU-Net model on our dataset of T2-weighted images demonstrate promising placental segmentation performance and indicate its potential for clinical application.