Scholarlink

<p>In clinical diagnosis, the automatic detection of anomalies using sophisticated deep learning (DL) architectures has become a highly sought-after and essential capability. This domain includes two distinct categories of pathology images: whole slide images (WSIs) and images consisting of individual patches. In state-of-the-art research, deep learning architectures have been validated using several public pathology datasets. Although the models demonstrated state-of-the-art performance on public datasets, further investigation is required to assess their performance on a private pathology dataset in practical cases. This study introduces a segmentation-based image quality analysis methodology that employs pretrained encoders (MobileNetV2, DenseNet121, VGG16, and ResNet50) with the baseline U-Net model. Before the training and validation phases, the images were preprocessed using the QuPath pathology tool, where they were resized and augmented to improve their suitability for the task. To evaluate the models, we used a private pathology dataset collected at the Samsung Medical Center of a Korean hospital, which consisted of whole slide images. It has three defects: bubble, tissue fold, and pen mark. Four different metrics were used to evaluate the performance of the models. The intersection over union (IoU) and highlight are introduced to gauge the performance of state-of-the-art models on real-life sample data. The experimental results, derived from a comprehensive analysis, revealed that the U-Net with the ResNet50 encoder model achieved an average accuracy of 98.17%, IoU of 98.07%, Dice score of 99.96%, and Jaccard distance of 0.0197 on the test split of all WSIs. Furthermore, we compared the model’s performance with that of the U-Net baseline architecture and pretrained architectures, wherein the U-Net with the ResNet50 encoder outperformed the U-Net model for the whole slide image dataset.</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

Deep segmentation-based anomaly detection from medical pathology images using a modified U-Net with pretrained deep encoders

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月25日

查看全文
  • Mahe Zabin,
  • Yechan Hwang,
  • Yuchae Jung,
  • Ho-Jin Choi

摘要

In clinical diagnosis, the automatic detection of anomalies using sophisticated deep learning (DL) architectures has become a highly sought-after and essential capability. This domain includes two distinct categories of pathology images: whole slide images (WSIs) and images consisting of individual patches. In state-of-the-art research, deep learning architectures have been validated using several public pathology datasets. Although the models demonstrated state-of-the-art performance on public datasets, further investigation is required to assess their performance on a private pathology dataset in practical cases. This study introduces a segmentation-based image quality analysis methodology that employs pretrained encoders (MobileNetV2, DenseNet121, VGG16, and ResNet50) with the baseline U-Net model. Before the training and validation phases, the images were preprocessed using the QuPath pathology tool, where they were resized and augmented to improve their suitability for the task. To evaluate the models, we used a private pathology dataset collected at the Samsung Medical Center of a Korean hospital, which consisted of whole slide images. It has three defects: bubble, tissue fold, and pen mark. Four different metrics were used to evaluate the performance of the models. The intersection over union (IoU) and highlight are introduced to gauge the performance of state-of-the-art models on real-life sample data. The experimental results, derived from a comprehensive analysis, revealed that the U-Net with the ResNet50 encoder model achieved an average accuracy of 98.17%, IoU of 98.07%, Dice score of 99.96%, and Jaccard distance of 0.0197 on the test split of all WSIs. Furthermore, we compared the model’s performance with that of the U-Net baseline architecture and pretrained architectures, wherein the U-Net with the ResNet50 encoder outperformed the U-Net model for the whole slide image dataset.