Scholarlink

<p>In the realm of cancer drug development, xenografted tumor models in mice are an invaluable tool for assessing efficacy and safety before one can continue testing in humans. CT scans are the most precise tool for analysis, but manual delineation is time-consuming and prone to bias, which calls for automatic segmentation tools to improve reproducibility and efficiency. We present TumSeg, a deep learning-based model for segmenting subcutaneous tumors in whole-body CT scans of mice. Developed and validated on a large, diverse dataset of 452 CT scans, TumSeg achieves a Dice score of 0.935 ± 0.036, outperforming previous methods. It accurately estimates tumor volume (1.5% ± 9.8% error), and extracts %ID<sub>mean</sub> (− 0.5% ± 4.4% error) and %ID<sub>max</sub> (0.1 ± 5.0% error) values from corresponding PET scans. TumSeg demonstrates robust performance across ten diverse datasets, handling variations in tumor size, location, and morphology. We also present an uncertainty quantification module that predicts segmentation quality, identifying all test set scans below a 0.850 Dice score without false positives. Additionally, TumSeg shows adaptability by successfully fine-tuning for MRI scans. With its high accuracy, speed, and ability to quantify uncertainty, TumSeg represents a significant advancement in preclinical tumor analysis.</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

Automatic deep learning-based tumor segmentation for PET-CT in mice with an uncertainty quantification module

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月27日

查看全文
  • Malte E. K. Jensen,
  • Andreas Clemmensen,
  • Jacob Gorm Hansen,
  • Julie van Krimpen Mortensen,
  • Emil S. Christensen,
  • Rasmus Sejersten Ripa,
  • Andreas Kjaer

摘要

In the realm of cancer drug development, xenografted tumor models in mice are an invaluable tool for assessing efficacy and safety before one can continue testing in humans. CT scans are the most precise tool for analysis, but manual delineation is time-consuming and prone to bias, which calls for automatic segmentation tools to improve reproducibility and efficiency. We present TumSeg, a deep learning-based model for segmenting subcutaneous tumors in whole-body CT scans of mice. Developed and validated on a large, diverse dataset of 452 CT scans, TumSeg achieves a Dice score of 0.935 ± 0.036, outperforming previous methods. It accurately estimates tumor volume (1.5% ± 9.8% error), and extracts %IDmean (− 0.5% ± 4.4% error) and %IDmax (0.1 ± 5.0% error) values from corresponding PET scans. TumSeg demonstrates robust performance across ten diverse datasets, handling variations in tumor size, location, and morphology. We also present an uncertainty quantification module that predicts segmentation quality, identifying all test set scans below a 0.850 Dice score without false positives. Additionally, TumSeg shows adaptability by successfully fine-tuning for MRI scans. With its high accuracy, speed, and ability to quantify uncertainty, TumSeg represents a significant advancement in preclinical tumor analysis.