Scholarlink

<p>Ovulation represents a fundamental physiological process in human reproduction, characterized by the release of an oocyte from the ovary. Infertility is a significant reproductive health condition that is frequently associated with ovarian dysfunction. The prevalence of infertility and the utilization of Assisted Reproductive Technologies (<i>ART</i>) has increased globally in recent decades. Trans-abdominal and transvaginal ultrasonography (<i>USG</i>) of the ovaries provide clinically relevant information on follicular count, size, spatial distribution, and their responsiveness to hormonal stimulation. Manual evaluation of large volumes of ultrasound images for follicle identification is both labor-intensive and susceptible to human error. This study proposes a new machine learning architecture, termed the widely integrated follicle extraction network (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(WIFE-Net\)</EquationSource> </InlineEquation>), which enables efficient and accurate follicle detection from ultrasound images. The model integrates both conventional and “Atrous” convolutional operations to enhance feature extraction from ultrasound (<i>USG</i>) images. Additionally, the model employs two “Pyramid Pooling Operation” to aggregate spatial information across multiple scales, enabling effective representation of object regions from coarse to fine levels. The proposed model facilitates fully automated follicle segmentation with high accuracy. A dataset comprising approximately 64000 annotated 2<i>D</i> ovarian ultrasound images, extracted from UltraSOund Volumes of Annotated ovaries in 3<i>D</i> (USOVA&#xa0;3D) image volumes, was used to train and evaluate the architecture. Experimental results demonstrate that the proposed model outperforms several state-of-the-art convolutional neural network (<i>CNN</i>)-based approaches on this dataset, highlighting its effectiveness and robustness in clinical applications. Specifically, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(WIFE-Net\)</EquationSource> </InlineEquation> achieved an accuracy of <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(97.78\%\)</EquationSource> </InlineEquation>, Jaccard Similarity Index of <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(75\%\)</EquationSource> </InlineEquation>, Recall value of <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(95.84\%\)</EquationSource> </InlineEquation>, F1 Score of <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(97\%\)</EquationSource> </InlineEquation>, precision of <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(97\%\)</EquationSource> </InlineEquation>, and Area Under<i>ROC</i> Curve (<i>AUC</i>) of <InlineEquation ID="IEq8"> <EquationSource Format="TEX">\(98\%\)</EquationSource> </InlineEquation>. These results indicate the potential applicability of the proposed method in clinical workflows for reliable follicle assessment.</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

WIFE-Net: widely integrated follicle extraction network

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Manas Sarkar,
  • Ardhendu Mandal,
  • Kanishka Sarkar

摘要

Ovulation represents a fundamental physiological process in human reproduction, characterized by the release of an oocyte from the ovary. Infertility is a significant reproductive health condition that is frequently associated with ovarian dysfunction. The prevalence of infertility and the utilization of Assisted Reproductive Technologies (ART) has increased globally in recent decades. Trans-abdominal and transvaginal ultrasonography (USG) of the ovaries provide clinically relevant information on follicular count, size, spatial distribution, and their responsiveness to hormonal stimulation. Manual evaluation of large volumes of ultrasound images for follicle identification is both labor-intensive and susceptible to human error. This study proposes a new machine learning architecture, termed the widely integrated follicle extraction network ( \(WIFE-Net\) ), which enables efficient and accurate follicle detection from ultrasound images. The model integrates both conventional and “Atrous” convolutional operations to enhance feature extraction from ultrasound (USG) images. Additionally, the model employs two “Pyramid Pooling Operation” to aggregate spatial information across multiple scales, enabling effective representation of object regions from coarse to fine levels. The proposed model facilitates fully automated follicle segmentation with high accuracy. A dataset comprising approximately 64000 annotated 2D ovarian ultrasound images, extracted from UltraSOund Volumes of Annotated ovaries in 3D (USOVA 3D) image volumes, was used to train and evaluate the architecture. Experimental results demonstrate that the proposed model outperforms several state-of-the-art convolutional neural network (CNN)-based approaches on this dataset, highlighting its effectiveness and robustness in clinical applications. Specifically, \(WIFE-Net\) achieved an accuracy of \(97.78\%\) , Jaccard Similarity Index of \(75\%\) , Recall value of \(95.84\%\) , F1 Score of \(97\%\) , precision of \(97\%\) , and Area UnderROC Curve (AUC) of \(98\%\) . These results indicate the potential applicability of the proposed method in clinical workflows for reliable follicle assessment.