Harvard researchers combine AI and imaging to reveal details conventional tissue tests miss

Researchers at Harvard Medical School have developed an artificial intelligence-assisted imaging method that can reveal details in tissue samples often missed by conventional techniques. The approach combines specialised microscopy with AI-powered image analysis to provide a more detailed picture of tissue structure and composition.

Named NoR-AI, the method captures information about proteins, lipids and water in tissue. It requires little to no pre-processing and can produce images in about 30 minutes, compared with the several days that traditional biopsy analysis can take, according to the researchers.

The findings were reported in a study published in the journal Proceedings of the National Academy of Sciences on November 5. Harvard Medical School published its account of the research on November 13.

The research addresses a limitation in conventional histopathology, the examination of tissue under a microscope to study disease. Traditional preparation methods can alter tissue structure and remove lipids and other water-soluble materials, leaving researchers without some information that could be relevant to understanding disease.

The new method is designed to preserve more of this information while producing images that can be analysed using artificial intelligence.

How the imaging method works

NoR-AI builds on a microscopy process called normalised Raman imaging, or NoRI, which uses stimulated Raman scattering to measure how light interacts with tissue. The technique then applies computational corrections to reduce the effects of light scattering and produce quantitative images of tissue components.

The researchers combined these images with a convolutional neural network, a type of AI system used to identify patterns in visual data. The network analyses the images to distinguish fine-scale features that may be difficult to identify using conventional tissue imaging.

With as few as two or three NoRI images of mouse kidneys, the AI system identified three distinct types of kidney tubules, classified anatomical regions and determined the sex of the animals based on the distribution of proteins and lipids.

Senior author Marc Kirschner, professor of systems biology at Harvard Medical School, said the technique allows researchers to measure the amounts of proteins, lipids and water in very small sections of tissue.

Co-first author William Trim, a research fellow in the Kirschner Lab, said the detailed images help the neural network identify distinctions that are difficult to detect in traditional histology images.

Studying changes linked to kidney injury

The team also tested the method using a mouse model of acute kidney injury, a condition in which kidney function is damaged. Working with Joseph Bonventre of Brigham and Women's Hospital, the researchers tracked changes in the organisation of proteins and lipids over 25 days following the injury.

The analysis helped the team measure known features of kidney disease and identify previously unrecognised changes associated with tissue damage and repair. These included variations in how lipids were lost and regained in different parts of the kidney over time.

The researchers said information about lipid distribution was particularly important in identifying these changes. Conventional tissue preparation often removes lipids, making them difficult to study through standard histological methods.

The work also demonstrated that NoR-AI could distinguish between different types of lipids, including cholesterol, saturated fats and unsaturated fats. This could help researchers investigate how these molecules are involved in processes such as inflammation and immune responses.

Potential applications beyond kidney research

Although the team used kidney tissue to demonstrate the method, researchers believe it could be applied to other organs and disease models. They are already using NoR-AI to study preclinical models of colon, liver and prostate cancers.

The team also plans to explore its application in diabetic kidney disease and other areas of research.

The method could eventually contribute to more detailed tissue evaluation and help researchers identify biological changes associated with ageing and disease. However, the findings so far are based on laboratory research involving mouse tissue. Further studies would be needed to establish how the approach performs in human samples and whether it can be used in routine clinical diagnosis.

The research was supported by the US National Institutes of Health. Kirschner and co-first author Seungeun Oh also hold a patent for the underlying Normalised Raman Imaging technology.