Artificial intelligence-based virtual staining platform for identifying tumor-associated macrophages from hematoxylin...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Biomarker discovery

Background:

• Virtual staining transforms H&E to IHC images using AI, offering a tissue-preserving alternative.
• Existing virtual staining methods struggle with image quality for accurate cell nuclei and IHC+ region delineation.
• VISTA is introduced as an AI platform to enhance H&E to virtual IHC translation.

Purpose of the Study:

• To develop and validate VISTA, an AI platform for high-quality virtual IHC generation from H&E images.
• To assess the prognostic significance of M2-TAM density identified through VISTA in HPV+ oropharyngeal squamous cell carcinoma.
• To compare VISTA's performance against existing virtual staining methods.

Main Methods:

• VISTA was applied to H&E images from 968 patients with HPV+ oropharyngeal squamous cell carcinoma.
• The platform was trained and tested using co-registered H&E and CD163+ IHC microarrays.
• M2-TAM density was calculated from VISTA-generated IHC images and evaluated for prognostic significance.

Main Results:

• High M2-TAM density correlated with worse overall survival in a validation cohort (p=0.0152, HR=1.63).
• VISTA produced superior virtual CD163+ IHC images compared to existing methods (SSIM=0.72, PSNR=21.5, FID=41.4).
• VISTA demonstrated enhanced performance in segmenting M2-TAMs (Dice=0.74).

Conclusions:

• VISTA is an effective computational platform for generating virtual IHC.
• The platform facilitates the discovery of novel biomarkers, such as M2-TAMs, directly from H&E images.
• Virtual IHC holds potential for advancing cancer research and diagnostics.