Related Experiment Video
Updated: Jun 14, 2025
Quantifying the Brain Metastatic Tumor Micro-Environment using an Organ-On-A Chip 3D Model, Machine Learning, and Confocal Tomography
Published on: August 16, 2020
Evaluation of tumor budding with virtual panCK stains generated by novel multi-model CNN framework
Xingzhong Hou1, Zhen Guan2, Xianwei Zhang3
1Institute of Computing Technology, Chinese Academy of Sciences, Beijing, 100190, China; School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing, 100190, China.
This study introduces a novel multi-model framework for virtual panCK staining, significantly improving cancer diagnosis accuracy and efficiency. The new method outperforms existing virtual staining techniques for evaluating tumor budding in breast cancer lymph nodes.
Area of Science:
- Computational pathology
- Artificial intelligence in medicine
- Digital pathology and image analysis
Background:
- Rising global cancer incidence necessitates faster, more accurate diagnoses.
- Hematoxylin and eosin (H&E)-pan-cytokeratin (panCK) stains are crucial for cancer diagnosis but chemical staining is time-consuming and irreversible.
- Virtual staining using generative adversarial networks (GANs) offers a promising alternative but current models face accuracy challenges.
Purpose of the Study:
- To develop a virtual panCK stain capable of replacing chemical panCK staining for cancer diagnosis.
- To address limitations in current virtual staining techniques for accurate stain generation.
- To evaluate the efficacy of virtual panCK stains in assessing tumor budding.
Main Methods:
- A multi-model framework combining Mask-RCNN for cell segmentation and GANs to extract cytokeratin distribution from H&E images.
- A tailored dynamic GAN model to convert H&E images into virtual panCK stains, incorporating cytokeratin distribution.
- Application of virtual panCK stains for tumor budding assessment in 45 H&E whole-slide images of breast cancer-invaded lymph nodes.
Main Results:
- The proposed multi-model virtual panCK stains demonstrated remarkable accuracy in tumor budding assessment, validated by pathologists and QuPath software.
- State-of-the-art single cycleGAN virtual panCK stains showed negligible accuracy in comparison.
- The framework efficiently generates dependable virtual panCK stains, reducing diagnostic turnaround times.
Conclusions:
- This study presents the first multi-model virtual panCK framework and its application in tumor budding assessment.
- The developed virtual panCK stains offer a highly accurate and efficient alternative to chemical staining, improving diagnostic speed and comprehension.
- The framework holds significant potential to advance virtual staining technology and enhance cancer diagnostic capabilities.
