Comparison of H&E staining-based methodologies for virtual stain generation in computational histopathology during production phase.
<p>Panel <b>A</b> illustrates the proposed unified H&E encoder approach, adapting the ComboGAN [<a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013516#pcbi.1013516.ref025" target="_blank">25</a>] approach to virtual st...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | , , , , |
| Published: |
2025
|
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | <p>Panel <b>A</b> illustrates the proposed unified H&E encoder approach, adapting the ComboGAN [<a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013516#pcbi.1013516.ref025" target="_blank">25</a>] approach to virtual staining, employing a single encoder and multiple decoders to generate various synthetic stains, thereby optimizing computational efficiency and scalability (to maintain focus on comparative methodology details on XAI capabilities are presented in <a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013516#pcbi.1013516.s001" target="_blank">S1 Fig</a>). Panel <b>B</b> depicts the traditional CycleGAN-like methodologies [<a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013516#pcbi.1013516.ref016" target="_blank">16</a>,<a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013516#pcbi.1013516.ref017" target="_blank">17</a>], which use multiple separate encoders and decoders for each stain, increasing the complexity of the model and computational demand. Panel <b>C</b> showcases the StarGAN-like approaches [<a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013516#pcbi.1013516.ref022" target="_blank">22</a>–<a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013516#pcbi.1013516.ref024" target="_blank">24</a>,<a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013516#pcbi.1013516.ref044" target="_blank">44</a>], using a style encoder and a single generator for multiple stains. While this architecture simplifies the model, it requires substantial computational resources and does not scale effectively, particularly as the number of stains increases (more stains, bigger generator), and still necessitates loading the large generator even for a subset of stains, leading to inefficiencies. The unified H&E approach in panel <b>A</b> represents a significant advance by reducing the need for multiple models and facilitating faster and more resource-efficient processing. This model is able to produce only the required stains, loading minimal model components into memory, reducing hardware requirements and computational costs in cloud-based deployments.</p> <p>(TIFF)</p> |
|---|