Deep learning creates virtual multiplexed immunostaining to improve cancer diagnosis

Researchers at the University of California, Los Angeles (UCLA), in collaboration with pathologists from Hadassah Hebrew University Medical Center and the University of Southern California, have developed a deep learning–based method that can digitally generate multiple immunohistochemical stains from a single, unstained tissue section.

The work is published in the journal BME Frontiers.

The approach enables accurate assessment of vascular invasion—a key indicator of cancer aggressiveness—without the need for conventional chemical staining procedures.

The study introduces a virtual multiplexed immunostaining (mIHC) framework that transforms autofluorescence microscopy images of label-free tissue into brightfield-equivalent images of hematoxylin and eosin (H&E) staining as well as two clinically important immunohistochemical markers: ERG, which labels endothelial cells, and PanCK, which highlights epithelial tumor cells. These virtual stains are generated simultaneously on the same physical tissue section using a single deep neural network.

Vascular invasion—defined as the presence of tumor cells within blood or lymphatic vessels—is a critical prognostic factor across many solid tumors, including thyroid cancer. However, its identification in routine pathology can be challenging. Traditional workflows rely on H&E slides followed by immunohistochemical staining on serial tissue sections, a process that is labor-intensive, costly, and prone to tissue loss and section-to-section variability.

UCLA’s virtual multiplexed staining approach allows pathologists to visualize tissue morphology, endothelial cells, and epithelial tumor cells on the same section, eliminating the risk of losing critical diagnostic information due to tissue dropout and enabling more reliable assessment of vascular invasion, said Aydogan Ozcan, Chancellor’s Professor of Electrical and Computer Engineering at UCLA and a corresponding senior author of the study.

The method is based on imaging unstained tissue using autofluorescence microscopy, followed by computational transformation using a conditional generative adversarial network. A digital staining matrix guides the network to generate different virtual stains from the same input image, ensuring precise spatial alignment between H&E and immunohistochemical outputs.

To validate the approach, the researchers applied the virtual mIHC framework to thyroid tissue microarrays and compared the digitally generated stains with conventional histochemical staining. Board-certified pathologists evaluated the images in a blinded study and found high concordance between virtual and traditional stains, with virtual staining often demonstrating superior consistency and specificity.

Because the stains are generated computationally, the results are highly reproducible and free from many of the artifacts seen in conventional immunohistochemistry, said Nir Pillar, a pathologist at Hadassah Hebrew University Medical Center and a corresponding author of the study. This has the potential to significantly improve diagnostic accuracy while reducing cost and turnaround time.

The virtual multiplexed immunostaining framework is compatible with existing digital pathology workflows and requires only a single tissue section.

After training, the system can generate virtual stains in seconds for individual fields of view and in minutes for whole-slide images, making it suitable for high-throughput clinical environments.

The researchers note that while the current study focused on thyroid cancer tissue, the approach is broadly applicable and could be extended to other tissue types and diagnostic markers. Future work will focus on large-scale, multi-center validation studies to further evaluate clinical performance and generalizability.