Scholarlink

<p>In IVF, it is very important to accurately and consistently assess the quality of embryos. This is because subjective evaluation can lead to bad transfer decisions. This work introduces an automated pipeline that identifies embryo regions from time-lapse micrographs and categorises day-3 and day-5 embryos as Good or Not Good, thereby minimising annotation effort and inter-observer variability. We suggest ResDense-GELU, a binary classification framework that uses a selectively fine-tuned ImageNet-pretrained ResNet18 and a dense feature refinement module to improve the separation of morphological features without using expensive attention mechanisms. Focal Loss mitigates class imbalance, and Stratified 5-Fold Cross-Validation with early stopping ensures that the model is robust. A comparative analysis of activation functions demonstrates that GELU consistently surpasses its alternatives. The best fold had 92.6% accuracy, 0.878 Macro F1, and 0.95 AUC. The overall cross-validation had a 0.8381 ± 0.0264 Macro F1 and a 0.8774 ± 0.0507 AUC. The model achieved about 87% accuracy on data it had never seen before, which shows that it can generalize well and be used in a clinical setting to objectively assess embryo quality.</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

Embryo quality classification using ResNet18 with dense feature refinement network for analysing microscopic bioimaging signals

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • A. Anushya,
  • Mohamed A. G. Hazber,
  • Hezam Saud Alrashidi

摘要

In IVF, it is very important to accurately and consistently assess the quality of embryos. This is because subjective evaluation can lead to bad transfer decisions. This work introduces an automated pipeline that identifies embryo regions from time-lapse micrographs and categorises day-3 and day-5 embryos as Good or Not Good, thereby minimising annotation effort and inter-observer variability. We suggest ResDense-GELU, a binary classification framework that uses a selectively fine-tuned ImageNet-pretrained ResNet18 and a dense feature refinement module to improve the separation of morphological features without using expensive attention mechanisms. Focal Loss mitigates class imbalance, and Stratified 5-Fold Cross-Validation with early stopping ensures that the model is robust. A comparative analysis of activation functions demonstrates that GELU consistently surpasses its alternatives. The best fold had 92.6% accuracy, 0.878 Macro F1, and 0.95 AUC. The overall cross-validation had a 0.8381 ± 0.0264 Macro F1 and a 0.8774 ± 0.0507 AUC. The model achieved about 87% accuracy on data it had never seen before, which shows that it can generalize well and be used in a clinical setting to objectively assess embryo quality.