2025 , Vol. 22 >Issue 08: 703 - 710
DOI: https://doi.org/10.3877/cma.j.issn.1672-6448.2025.08.004
基于注意力机制改进的子宫解剖结构检测与分割多任务模型的性能评估
通信作者:
李胜利,Email:lishengli63@126.comCopy editor: 汪荣
收稿日期: 2025-06-20
网络出版日期: 2025-09-29
基金资助
深圳市自然科学基金基础研究面上项目(JCYJ20240813115114020)
版权
Performance of an attention-enhanced multi-task model for uterine anatomical structure detection and segmentation
Corresponding author:
Li Shengli, Email: lishengli63@126.comReceived date: 2025-06-20
Online published: 2025-09-29
Copyright
以YOLOv8框架为基础,引入高效多尺度注意力(EMA)机制,构建子宫解剖结构检测与分割智能模型(IMSU),并评估其性能。
回顾性收集2021年1月至2022年12月深圳市妇幼保健院共计4326张非妊娠子宫正中矢状切面超声图像,人工标注子宫宫体、宫颈及内膜3个关键解剖结构,建立图像数据库。按8∶1∶1比例划分为训练集(3460张)、验证集(433张)与测试集(424张),引入EMA机制对YOLOv8模型进行改进,构建IMSU。首先训练并验证原始YOLOv8标准模型及基于EMA模块改进的模型(IMSU),随后在测试集上对两种模型对子宫宫体、宫颈及内膜的自动结构检测与分割性能进行评估,评估指标包括精确率、召回率、平均精度(mAP)2个层级指标(mAP@50与mAP@50-95)。
在3个关键结构自动检测任务中,IMSU的整体平均精确率(0.920 vs 0.905)、召回率(0.939 vs 0.917)及mAP@50(0.952 vs 0.933)均优于YOLOv8;尤其对宫颈的检测mAP@50值由0.858提升至0.919,召回率由0.778提升至0.842。在自动分割任务中,与YOLOv8相比,IMSU整体平均精确率由0.905提升至0.914,召回率由0.915提升至0.933,mAP@50由0.930提升至0.952,mAP@50-95亦从0.661提升至0.677;对宫颈的分割mAP@50-95由0.570提升至0.597。
融合EMA机制的IMSU显著提升了非妊娠子宫正中矢状切面关键结构的自动检测性能与分割精度,为实现子宫结构的智能量化测量及超声辅助诊断提供了技术支持,具有良好的临床应用前景。
江瑶 , 蒋程 , 余翔 , 谭莹 , 温昕 , 温慧莹 , 彭桂艳 , 李胜利 . 基于注意力机制改进的子宫解剖结构检测与分割多任务模型的性能评估[J]. 中华医学超声杂志(电子版), 2025 , 22(08) : 703 -710 . DOI: 10.3877/cma.j.issn.1672-6448.2025.08.004
To develop an intelligent model for uterine anatomical structure detection and segmentation (IMSU) by integrating the Efficient Multi-scale Attention (EMA) mechanism into the You Only Look Once version 8 (YOLOv8) framework and evaluate its performance.
A total of 4326 non-pregnant mid-sagittal uterine ultrasound images were retrospectively collected from Shenzhen Maternity and Child Healthcare Hospital (January 2021-December 2022). Three key anatomical structures (uterine corpus, cervix, and endometrium) were manually annotated to establish an image database. The dataset was divided into training (3460 images), validation (433 images), and test sets (424 images) at an 8∶1∶1 ratio. An IMSU model was constructed by enhancing YOLOv8 with the EMA module. Both the baseline YOLOv8 and IMSU models were trained and validated, followed by performance evaluation on the test set for automated detection and segmentation of uterine structures. Metrics included precision, recall, and mean Average Precision (mAP) at two levels: mAP@50 and mAP@50-95.
For detection tasks, IMSU outperformed YOLOv8 in overall precision (0.920 vs 0.905), recall (0.939 vs 0.917), and mAP@50 (0.952 vs 0.933). Notably, cervical detection mAP@50 improved from 0.858 to 0.919 and recall increased from 0.778 to 0.842. In segmentation tasks, IMSU achieved higher precision (0.914 vs 0.905), recall (0.933 vs 0.915), mAP@50 (0.952 vs 0.930), and mAP@50-95 (0.677 vs 0.661). Cervical segmentation mAP@50-95 rose from 0.570 to 0.597.
The EMA-enhanced IMSU significantly improves automated detection and segmentation accuracy for key uterine structures in mid-sagittal ultrasound images, providing technical support for intelligent quantitative uterine measurements and ultrasound-assisted diagnosis with promising clinical applicability.
图2 高效多尺度注意力(EMA)机制的结构示意图。输入数据首先被分组处理,每组数据尺寸为C/G×H×W,其中C、H、W分别代表通道数、高度和宽度,G为分组数;在分组内部,数据通过X轴池化和Y轴池化进行特征提取,并经过拼接与1×1卷积层处理,随后通过Sigmoid激活函数生成权重;同时,3×3卷积层处理后的特征通过Softmax激活函数计算得到归一化的权重;这些权重与原始特征通过矩阵乘法结合,再与另一路径的直接矩阵乘法结果相加,最终通过Sigmoid激活函数输出调整后的特征图,尺寸保持为C×H×W;整个过程体现了EMA机制对不同特征通道和空间位置的动态加权,以增强模型对关键信息的关注度 |


表1 YOLOv8标准模型与改进的IMSU对非妊娠期子宫正中矢状切面结构的检测结果 |
| 解剖结构 | 检测精确率 | 检测召回率 | 检测mAP@50 | 检测mAP@50-95 | ||||
|---|---|---|---|---|---|---|---|---|
| IMSU | YOLOv8 | IMSU | YOLOv8 | IMSU | YOLOv8 | IMSU | YOLOv8 | |
| 子宫宫体 | 0.980 | 0.969 | 0.992 | 0.992 | 0.991 | 0.994 | 0.912 | 0.884 |
| 子宫内膜 | 0.939 | 0.914 | 0.983 | 0.983 | 0.946 | 0.948 | 0.611 | 0.600 |
| 子宫宫颈 | 0.840 | 0.831 | 0.842 | 0.778 | 0.919 | 0.858 | 0.654 | 0.617 |
| 整体平均 | 0.920 | 0.905 | 0.939 | 0.917 | 0.952 | 0.933 | 0.726 | 0.700 |
注:IMSU为基于YOLOv8改进的非妊娠期子宫正中矢状切面智能模型;mAP@50为交并比阈值为0.5时各类别平均精度;mAP@50-95为交并比阈值从0.5递增至0.95之间的平均精度 |
表2 YOLOv8标准模型与改进的IMSU对非妊娠期子宫正中矢状切面结构的分割结果 |
| 解剖结构 | 分割精确率 | 分割召回率 | 分割mAP@50 | 分割mAP@50-95 | ||||
|---|---|---|---|---|---|---|---|---|
| IMSU | YOLOv8 | IMSU | YOLOv8 | IMSU | YOLOv8 | IMSU | YOLOv8 | |
| 子宫宫体 | 0.980 | 0.980 | 0.992 | 0.992 | 0.991 | 0.994 | 0.897 | 0.895 |
| 子宫内膜 | 0.923 | 0.903 | 0.965 | 0.973 | 0.946 | 0.939 | 0.535 | 0.519 |
| 子宫宫颈 | 0.840 | 0.832 | 0.842 | 0.781 | 0.919 | 0.858 | 0.597 | 0.570 |
| 整体平均 | 0.914 | 0.905 | 0.933 | 0.915 | 0.952 | 0.930 | 0.677 | 0.661 |
注:IMSU为基于YOLOv8改进的非妊娠期子宫正中矢状切面智能模型;mAP@50为交并比阈值为0.5时各类别平均精度;mAP@50-95为交并比阈值从0.5递增至0.95之间的平均精度 |
| 1 |
AIUM practice parameter for the performance of ultrasound of the female pelvis, 2024 revision [J]. J Ultrasound Med, 2024, 43(11): E56-E64.
|
| 2 |
|
| 3 |
|
| 4 |
|
| 5 |
|
| 6 |
|
| 7 |
|
| 8 |
张朝阳, 张上, 胡益民, 等. 动态聚焦多维注意力遥感弱小目标检测[J]. 无线电通信技术, 2025, 51(1): 196-209.
|
| 9 |
苏警. 基于深度学习的图像分类与目标检测算法研究[J]. 电脑编程技巧与维护, 2025, (5): 131-133.
|
| 10 |
杨仕伟, 南新元, 蔡鑫, 等. PAMNS-Net: 基于原型学习和自适应多分支融合的细胞核分割网络[J]. 激光与光电子学进展, 2025, 62(24): 3.
|
| 11 |
|
| 12 |
|
| 13 |
林小辉, 陈晓华. 阴道超声测量宫颈长度及子宫宫颈前角用于早产预测中的价值[J]. 中国医疗器械信息, 2024, 36(3): 36-38.
|
| 14 |
林举, 陈德高, 林玉玲, 等. 阴道超声测量子宫内膜厚度与异常子宫出血者组织病理学结果的相关性[J]. 中国卫生工程学, 2022, 21(4): 661-663.
|
| 15 |
|
/
| 〈 |
|
〉 |