Inside the Algorithms: How AI Is Reshaping Pathology, From Digital Slides to Bias-Aware Diagnosis

When a patient needs a second opinion on a biopsy, the tissue often travels by hand. Glass slides carrying biological material are carried from one office to another so that more than one pathologist can look at them. That is a surprisingly analog routine for 2026, and it is the problem that digital pathology is meant to solve.

Lee Cooper, director of computational pathology at Northwestern University's Feinberg School of Medicine, says digital imaging is becoming more common in the field. A sample of blood or tissue is placed on a glass slide, the slide goes into a scanner, and the scanner produces a digital image that can be shared electronically. Once a sample is digital, algorithms can review it too. Cooper runs a lab where researchers build those algorithms, and he expects them to help pathologists reach more accurate and more detailed diagnoses, and to help researchers understand how disease progressed in past patients. The advantage is scale: a single image can hold a million cells or more, and Cooper notes that measuring that precisely is very hard for a person.

Inside the Cooper Lab, researchers write algorithms, digitize specimens and analyze the resulting images, drawing on samples from ongoing studies as well as banked ones. Each person works on a particular data set or a specific problem pathologists face. Fattaneh Pourakpour, a Ph.D. student in electrical engineering, uses tissue from the Mesulam Institute for Cognitive Neurology and Alzheimer's Disease brain bank, which accepts brain donations from people who had dementia. Her segmentation algorithms separate regions such as gray and white matter, and her cell classifiers identify and categorize cell types within the gray matter. She is looking for biomarkers tied to Alzheimer's and other neurodegenerative disorders.

Post-doctoral researcher Ramin Nateghi, who holds a Ph.D. from Shiraz University of Technology in Iran, applies machine learning to prostate cancer. The disease often moves slowly, and he says biopsies taken over 25 years may show few visible differences. He hopes the tools can detect subtle changes that the human eye misses and help doctors make better decisions.

None of this is meant to replace pathologists. Cooper points to edge cases: the models rely on correlation, with no causal understanding inside them, so they are limited by their training data and less useful on a rare case they have never seen. Whether such models understand their subject deeply or only superficially remains an open question. Cooper still finds them useful for accurate, efficient predictions, since they can weigh far more properties of a sample than a doctor can, but he does not think they should be trusted implicitly.

The need is real. Cooper says pathology is by far the largest source of data in medicine, and the Northwestern Medicine system alone generates about 2 million glass slides a year. Lawrence Chillrud, a Ph.D. student in electrical engineering, describes a shortage of pathologists in the United States set against heaps of biopsy data awaiting review. His models run a preliminary screen, setting aside the most obvious cases and flagging others for fast-tracked specialist review. Similar tools could spread the hardest cases across a group of pathologists or move them early in the workflow, count particular cell types, point attention to important areas of a sample, or check a pathologist's report for errors.

Chillrud is also building uncertainty tools that estimate how likely a model's prediction is to be right. Many models, he says, are overconfident and give a false picture of how far they can be trusted. Such a tool would warn the pathologist when an error is probable. It could also limit bias: a model trained mostly on white, affluent patients from one part of a city, who tend to come to a given hospital, will predict less well for patients from other backgrounds, and an uncertainty flag can signal those gaps. The tools might also help smaller and rural systems, where specialists, such as someone qualified to give a detailed prostate biopsy read, are harder to find than in large city systems.

Cooper is unsure how fast adoption will come. Tools need FDA clearance and validation on very high-quality data that is hard to obtain, and not every health system has the scanning technology Northwestern does. The data load is heavy too: each image holds several billion pixels, and Northwestern's pathology would produce about 3 petabytes, or 3 million gigabytes, a year. Still, he is encouraged that scientists, not just tool developers, are using the tools to relate tissue patterns to outcomes such as treatment success. What to watch is whether clearance, validated data and wider digitization reach smaller hospitals.