Scholarlink

<p>Accurate assessment of fetal development is a critical component of prenatal care, particularly the identification of external morphological deformities. Traditionally, this process relies heavily on manual interpretation of 2D ultrasound images by trained sonographers, which is time-consuming and often subject to significant inter-observer variability. To address these limitations, this study presents a fully automated deep learning based framework for the segmentation and identification of key fetal anatomical structures from 2D ultrasound scans. Specifically, the proposed model targets four primary fetal body regions: the arms, legs, head, and abdomen. Modified attention U-Net architecture is employed for segmentation due to its capability to focus on relevant spatial features while suppressing background noise typical in ultrasound imaging. The model is trained over a period of ten epochs for each anatomical region. At the final epoch, the model achieved high segmentation accuracy across all regions: 99.33% for the abdomen, 99.33% for the head, 99.45% for the legs, and 99.38% for the arms. Corresponding loss values are recorded as 0.0150, 0.0149, 0.0125, and 0.0140, respectively. These metrics indicate robust convergence behavior and minimal overfitting, reflecting the model’s ability to generalize effectively across variations in fetal pose and image orientation.</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

Automated detection and segmentation of fetal body parts in 2D ultrasound images using modified U-Net architecture for external deformity assessment

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Rupashree Mohanty,
  • Santosh Kumar Pani,
  • Satya Ranjan Dash

摘要

Accurate assessment of fetal development is a critical component of prenatal care, particularly the identification of external morphological deformities. Traditionally, this process relies heavily on manual interpretation of 2D ultrasound images by trained sonographers, which is time-consuming and often subject to significant inter-observer variability. To address these limitations, this study presents a fully automated deep learning based framework for the segmentation and identification of key fetal anatomical structures from 2D ultrasound scans. Specifically, the proposed model targets four primary fetal body regions: the arms, legs, head, and abdomen. Modified attention U-Net architecture is employed for segmentation due to its capability to focus on relevant spatial features while suppressing background noise typical in ultrasound imaging. The model is trained over a period of ten epochs for each anatomical region. At the final epoch, the model achieved high segmentation accuracy across all regions: 99.33% for the abdomen, 99.33% for the head, 99.45% for the legs, and 99.38% for the arms. Corresponding loss values are recorded as 0.0150, 0.0149, 0.0125, and 0.0140, respectively. These metrics indicate robust convergence behavior and minimal overfitting, reflecting the model’s ability to generalize effectively across variations in fetal pose and image orientation.