MADRNet:一种形态意识的双路径可逆网络,用于精子分类
IEEE journal of biomedical and health informatics
|November 24, 2025
概括
这项研究介绍了MADRNet,这是一种用于准确分析精子形态的AI模型,改进了男性不孕症诊断. 该模型实现了高精度和实时处理,解决了手动评估和现有的自动化方法的局限性.
科学领域:
- 生殖生物学 生殖生物学
- 医疗成像医学成像
- 人工智能的人工智能
背景情况:
- 手动的精子形态分析是主观的,影响男性不孕症诊断的准确性.
- 现有的自动化精子分类模型通常需要人工干预,并且与世卫组织标准不一致.
- 临床评估依赖于生物标志物,如头部面积比和体完整性.
研究的目的:
- 开发一个自动化,准确和高效的精子形态分析模型,符合世卫组织的标准.
- 克服手动精子评估的主观性和不一致性.
- 将关键的临床生物标志物整合到人工智能框架中,以改善男性不孕症诊断.
主要方法:
- 提出了一个具有形态意识的双路径可逆网络 (MADRNet),结合了头部面积比和体完整性.
- 使用双路径注意力机制 (空间和通道注意力) 与宏观体解剖学约束.
- 开发了一个动态损失函数,用于头角比限制,并采用可逆架构.
- 在HuSHeM数据集上训练和评估模型.
主要成果:
- 在HuSHeM数据集上,MADRNet实现了96.3%的准确性和96.8%的F1得分.
- 该模型以每张图像32毫秒的速度进行实时处理.
- 可逆架构保留了微观细节,同时减少了GPU内存使用量.
结论:
- MADRNet为临床精子查提供了精确高效的自动化解决方案.
- 该模型有效地解决了手动分析和现有的自动化方法的局限性.
- 开发的AI工具显示了改善男性不孕症诊断的巨大潜力.
相关概念视频
Sperm Structure and Semen Composition
10.1K
During ejaculation, males release around 2-5 milliliters of semen, which is a complex mixture of mature sperm and various fluids produced by accessory glands. The mature sperm cells measure approximately 60 micrometers in length and consist of a head, neck, midpiece, and tail. The head is flattened and tapered, measuring about 4 to 5 micrometers in length. It contains a nucleus with condensed chromosomes and an acrosome, a cap-like structure filled with enzymes essential for penetrating the...
10.1K
Spermatogenesis
121.4K
Spermatogenesis is the process by which haploid sperm cells are produced in the male testes. It starts with stem cells located close to the outer rim of seminiferous tubules. These spermatogonial stem cells divide asymmetrically to give rise to additional stem cells (meaning that these structures “self-renew”), as well as sperm progenitors, called spermatocytes. Importantly, this method of asymmetric mitotic division maintains a population of spermatogonial stem cells in the male...
121.4K


