Researchers from the University of Eastern Finland, the University of Turku, and Tampere University have developed a synthetic intelligence-based method for virtual staining of histopathological tissue samples as a component of the Nordic ABCAP consortium. Chemical staining has been the cornerstone of studying histopathology for greater than a century and is widely applied in, for instance, cancer diagnostics. 

An example of virtual staining of tissue. Unstained tissue on the left, chemically stained tissue in the center and virtually stained tissue on the fitting. The examples are prostate tissue. Image Credit: University of Turku

“Chemical staining makes the morphology of the virtually transparent, low-contrast tissue sections visible. Without it, analyzing tissue morphology is sort of unattainable for human vision. Chemical staining is irreversible, and generally, it prevents the usage of the identical sample for other experiments or measurements,” says University Researcher and Vice Director of the Institute of Biomedicine on the University of Eastern Finland Leena Latonen, who led the experimental a part of the study.

The bogus intelligence method developed on this study produces computational images that very closely resemble those produced by the actual chemical staining process. This virtually stained image can then be used for inspecting the morphology of the tissues. Virtual staining reduces each the chemical burden and manual work needed for sample processing while also enabling the usage of the tissue for other purposes than the staining itself.

The strength of the proposed virtual staining method is that it requires no special hardware or infrastructure beyond a daily light microscopy and an appropriate computer.

“The outcomes are very widely applicable. There are many topics for follow-up research, and the computational methods can still be improved. Nonetheless, we will already envision several application areas where virtual staining can have a serious impact in histopathology,” says Associate Professor Pekka Ruusuvuori from the University of Turku, who led the computational a part of the study.

Ground-breaking research with international funding

One in every of the important thing aspects enabling the study was the consortium funding obtained from the ERAPerMed joint transnational call. The ABCAP consortium consists of Nordic research groups developing artificial intelligence-based diagnostics of breast cancer towards personalized medicine and is funded by ERAPerMed, Nordic Cancer Union and the Academy of Finland. Each Latonen and Ruusuvuori lead their very own subprojects.

“This research is actually cross-disciplinary. Without consortium funding, it could be very difficult to search out enough resources for each the experimental laboratory work and the computational effort to enable studies like this,” acknowledge Ruusuvuori and Latonen.

This cross-disciplinary research is predicated on expertise in tissue biology, histological processes, bioimage informatics and artificial intelligence. The primary a part of the two-phase study focused on optimizing the tissue sample processing and imaging steps, and was carried out by Doctoral Researcher Sonja Koivukoski from the University of Eastern Finland. Systematic assessment of histological feasibility was a novel component within the study.

“Development of computational methods using artificial intelligence often lacks proper assessment of the feasibility from the attitude of the tip user. This will result in methods being developed and published but eventually not likely utilized in practice. Due to this fact, it is very vital to mix each computational and domain-based knowledge already in the event phase, as was done in our study,” state Latonen and Koivukoski.

Great potential of computational methods

Deep neural networks learning form large volumes of information have rapidly transformed the sphere of biomedical image evaluation. Along with traditional image evaluation tasks, equivalent to image interpretation, these methods are also well fitted to image-to-image transforms. Virtual staining is an example of such a task, as was successfully shown within the two published parts of the work. The second part focused on optimizing virtual staining based on generative adversarial neural networks, with Doctoral Researcher Umair Khan from the University of Turku because the lead developer.

Deep neural networks are able to acting at a level we weren’t capable of imagine some time ago. Artificial intelligence-based virtual staining can have a serious impact towards more efficient sample processing in histopathology.”

Umair Khan, Doctoral Researcher, University of Turku

Along with the unreal intelligence algorithms, the important thing to success was the supply of high-performance computing services through CSC.

“In Finland, we’ve got a wonderful infrastructure for parallel high-performance computing. Computationally intensive research like this might not be possible without the capability provided by CSC,” says Ruusuvuori.

The outcomes of the study were published in two international peer-reviewed journals, Laboratory Investigation and Patterns.

Source:

Journal references:

• Koivukoski, S., et al. (2023). Unstained tissue imaging and virtual hematoxylin and eosin staining of histological whole slide images. Laboratory Investigation. doi.org/10.1016/j.labinv.2023.100070
• Khan, U., et al. The effect of neural network architecture on virtual H&E staining: Systematic assessment of histological feasibility. Patterns. doi.org/10.1016/j.patter.2023.100725