Deep Learning–Based Method Improves Cancer Diagnosis

Conventional methods require multiple chemical stains to be applied to separate tissue sections, which increases cost, processing time, and the risk of losing diagnostic information. Researchers have now developed an artificial intelligence (AI)-based approach that can digitally generate multiple diagnostic stains from a single unstained tissue section, enabling more accurate assessment of cancer invasion.

Researchers at the University of California, Los Angeles (UCLA, Los Angeles, CA, USA), in collaboration with Hadassah Hebrew University Medical Center (Jerusalem, Israel) and the University of Southern California (USC, Los Angeles, CA, USA), have introduced a virtual multiplexed immunohistochemistry framework that converts autofluorescence microscopy images of label-free tissue into brightfield-equivalent images of hematoxylin and eosin staining, along with two key immunohistochemical markers: ERG for endothelial cells and PanCK for epithelial tumor cells.

Unstained tissue sections were first imaged using autofluorescence microscopy. A conditional generative adversarial network then digitally transformed these images into multiple virtual stains using a single deep neural network. A digital staining matrix guided the model to produce precisely aligned virtual H&E, ERG, and PanCK images from the same tissue section, eliminating the need for serial sectioning and chemical staining.

When applied to thyroid tissue microarrays, the virtual staining approach showed high concordance with conventional histochemical and immunohistochemical stains. In blinded evaluations, board-certified pathologists found that the digitally generated stains were comparable to traditional methods and, in some cases, demonstrated improved consistency and specificity. The approach, presented in BME Frontiers, enabled clearer visualization of tumor cells within blood or lymphatic vessels, supporting more reliable identification of vascular invasion.

Because the staining is generated computationally, the method avoids many artifacts associated with conventional immunohistochemistry and delivers highly reproducible results. Virtual stains can be produced within seconds for individual regions and in minutes for whole-slide images, making the framework compatible with high-throughput digital pathology workflows. While demonstrated in thyroid cancer, the approach could be extended to other tumor types and additional diagnostic markers, pending further multi-center clinical validation.

Related Links:
UCLA
Hadassah Hebrew University Medical Center
USC