Scientists based at the University of Eastern Finland, together with colleagues from the University of Turku, and Tampere University, have developed an artificial intelligence-based method for virtual staining of histopathological tissue samples. This collaboration has taken place through the Nordic ABCAP consortium.
Chemical staining is the cornerstone of studying histopathology (the diagnosis and study of diseases of the tissues) and it is essential for medical understanding, such as for cancer diagnostics. A trained histopathologist is able to view potentially cancerous or atypical tissues and aid other medical specialists in making diagnoses or assessing the effectiveness of treatments.
Good quality staining enables the morphology of what would otherwise be almost transparent, low-contrast tissue sections to become visible. However, a limitation with the approach is that chemical staining is irreversible, and it prevents the use of the same sample for other experiments or measurements.
An alternative approach is by deploying artificial intelligence. The new computational method generates computational images that very closely resemble those produced by the actual chemical staining process.
The virtual staining concept not only seeks to improve accuracy it also automates part of the laboratory operation by reducing the manual work required to process samples. The virtual staining aspect was achieved through image-to-image transforms.
In terms of equipment, the process can be achieved using standard infrastructure: Regular light microscopy and a suitable computer equipped with the AI program.
The AI was developed by different researchers, enabling cross-disciplinary practices to embedded, including tissue biology, histological processes, bioimage informatics and artificial intelligence.
The development was carried out in different phases, beginning with optimising the tissue sample processing and imaging steps. This required a systematic assessment of histological feasibility was a unique component in the study.
The next phase involved training the deep neural network using large volumes of data. Optimising the virtual staining was based on generative adversarial neural networks. This refers to a machine learning model in which two neural networks compete with each other by using deep learning methods to become more accurate in their predictions.
The research is described in two journals: Laboratory Investigation ‘Unstained tissue imaging and virtual hematoxylin and eosin staining of histological whole slide images’; and Patterns ‘The effect of neural network architecture on virtual H&E staining: Systematic assessment of histological feasibility’.
