Harvard researchers combine AI and imaging to reveal kidney-injury changes in mice

Harvard Medical School researchers have combined a specialized microscope technique with AI to reveal fine-scale changes in injured mouse kidneys. Their pipeline, NoR-AI, captures fats that conventional tissue preparation typically removes, alongside proteins and water. The team used it to track previously unobserved patterns of fat loss and gain as kidney damage and repair unfolded.

Harvard Medical School described the findings on November 13, pointing to a study published November 5 in Proceedings of the National Academy of Sciences. The researchers say the imaging pipeline can be completed in 30 minutes, compared with several days for traditional biopsy analysis. The kidney work serves as a proof of principle.

Keeping information before asking AI to find it

Histopathology—the study of disease in tissue under a microscope—depends on how samples are prepared. Traditional preparation can distort subtle structures and remove lipids, or fats, along with water-soluble materials. That leaves researchers without some chemical information that may be relevant to disease. NoR-AI requires little to no tissue pre-processing and produces quantitative images of lipid, protein and water content.

The pipeline starts with normalized Raman imaging, known as NoRI, developed by co-first author Seungeun Oh while working in Marc Kirschner’s lab. It uses stimulated Raman scattering, which measures changes in light as it interacts with tissue. Software then corrects for light-scattering effects that can interfere with the measurements. The result supplies the image-analysis stage with clearer, quantitative information.

A convolutional neural network, a type of AI used to analyze images, then searches those measurements for fine distinctions. Co-first author William Trim said traditional histology images force neural networks to work through substantial noise. The researchers’ approach changes both what the microscope captures and what the AI can distinguish—not just the algorithm reading an existing slide.

You wouldn’t have the lipid data without NoRI, and you wouldn’t spot these tiny changes without the neural network

William Trim, co-first author and research fellow in systems biology at Harvard Medical School

Following kidney damage across 25 days

With as few as two or three NoRI images of mouse kidneys, the neural network performed several classification tasks. It could distinguish tissue structures and extract information from the distribution of lipids and proteins. The team reported that it could:

• Identify three distinct types of kidney tubules.
• Classify anatomical regions within the kidney.
• Determine the mouse’s sex from lipid and protein distributions in the tubules.

The injury experiment went beyond identifying normal anatomy. Working with Joseph Bonventre at Brigham and Women’s Hospital, the researchers tracked lipid and protein organization over 25 days after acute kidney injury in mice. They quantified several known features of kidney disease and identified previously unseen aspects of damage and repair, including where lipids disappeared and reappeared over time.

A research tool moving beyond the kidney

The researchers also say NoR-AI can distinguish cholesterol, saturated fats and unsaturated fats. That could make it useful for studying how different lipids contribute to inflammation and immune responses. Improved clinical evaluation is a potential application, rather than an outcome demonstrated by this mouse-kidney proof of principle.

The National Institutes of Health supported the study. Harvard’s disclosure notes that Kirschner and Oh hold a patent for normalized Raman imaging.

The team is already using NoR-AI in preclinical models of colon, liver and prostate cancers. Diabetic kidney disease is among its planned next applications, extending the work from tracking acute injury toward other disease settings.