Biomedical Research · us
Seeing the Lipids Lost During Tissue Section Preparation: Harvard Combines Raman Imaging and AI to Track Kidney Injury and Repair
A mouse study turns proteins, lipids, and water into quantitative tissue images, capturing subtle changes after kidney injury. This opens a new window into disease progression but cannot yet be considered a diagnostic tool for humans.
After kidney injury, how do lipids within cells disappear, accumulate, and redistribute during repair? Conventional tissue sections can reveal many pathological changes, but preparation may remove lipids, leaving some clues absent under the microscope. A Harvard Medical School research team combined Raman microscopy with artificial intelligence to quantify these easily overlooked changes in mouse kidneys, adding another layer of information to help understand injury and repair.
The workflow, called NoR-AI, pairs normalized Raman imaging (NoRI) with convolutional neural networks. The former uses stimulated Raman scattering to obtain molecular signals, then computationally corrects interference caused by light scattering to quantify proteins, lipids, and water in tissue; the latter identifies structures and subtle differences in the images. The study was published online on November 5 in the Proceedings of the National Academy of Sciences (PNAS).
In mouse kidney images, the model achieved F1 scores of 0.93, 0.91, and 0.97 for identifying renal tubule types, anatomical regions, and biological sex, respectively. F1 combines precision and recall; a score closer to 1 indicates better performance on the classification task. These figures reflect image recognition capabilities under the study conditions and cannot be directly translated into diagnostic accuracy for human diseases.
The team further applied the workflow to an acute kidney injury model, tracking protein and lipid distributions over the 25 days following injury. They quantified remodeling of the renal tubular brush border and changes in intracellular lipid droplets. The brush border consists of densely packed microvilli on the surface of renal tubular cells. Placing these structures and lipid changes on the same quantitative map allowed researchers to compare patterns of injury and repair across locations and time points.
Quantitative correction also produced measurable differences. When distinguishing between early time points after ischemia-reperfusion injury, 24 to 48 hours after injury, NoRI images achieved a classification F1 score of 0.99, compared with 0.84 for raw stimulated Raman scattering images. This comparison supports the value of normalization, but it compares two forms of Raman imaging rather than directly testing performance against pathologists interpreting conventionally stained tissue sections.
Speed and sample processing offer another potential advantage. According to a Harvard Medical School report, the workflow requires little or no tissue preprocessing and can be completed in about 30 minutes. Medical Xpress republished the same institutional report on November 16, noting that the team is applying the method to preclinical models of colon, liver, and prostate cancers; these extensions remain exploratory research.
The strongest finding so far is the preservation and quantification of information in mouse tissue that is otherwise easily lost, allowing kidney disease progression to be studied in finer detail. Clinical use will still require validation in human specimens, across different pathological conditions, and under conditions in different laboratories. It will also be necessary to determine whether the additional imaging features can improve diagnostic or treatment decisions. Seeing more tissue detail is already a research advance; its clinical significance requires further evidence.