For pathologists managing acute myeloid leukemia (AML), evaluating TP53 status often means choosing between time-consuming manual cell counting and expensive genetic testing. But that might be about to change, thanks to work from researchers at the University of Toronto.

Fatemeh Zabihollahy, PhD, lead AI research scientist at the University of Toronto, presented findings that could change how pathology labs assess TP53 mutations in patients with AML.[1] Her team has developed an AI system that can analyze immunohistochemistry (IHC) slides in a fraction of the time it takes a pathologist without sacrificing accuracy.

Approximately half of patients with leukemia carry TP53 mutations, which are associated with poor prognosis and should be considered when choosing the appropriate treatment for patients. Although next-generation sequencing (NGS) can be used to identify these mutations, it has significant drawbacks. “NGS is still our gold standard, but at our institution, it takes about two weeks to get results back,” said Zabihollahy. “Not to mention the cost — many centers can’t offer it routinely.”

IHC, on the other hand, is widely available and relatively inexpensive; however, manual interpretation of IHC slides is time consuming. Pathologists must examine multiple fields and distinguish between negative cells (which appear blue) and positive cells (which show brown staining of varying intensity). “Right now, pathologists have to analyze at least 3,000 cells per case,” explained Zabihollahy. “That’s 10 to 15 minutes of careful counting for each patient, looking not just for positive cells but also assessing their staining intensity.”

To address this challenge, the team developed an AI method for automated cell detection. Instead of using traditional bounding boxes to identify cells, their AI system employs a StarDist-based model, which is a state-of-the-art method for cell segmentation. Zabihollahy explained that StarDist can better handle overlapping cells. “We needed to solve the overlap problem,” said Zabihollahy. “When cells overlap, conventional AI methods tend to count them as one cell, which throws off our numbers. Our approach uses star-convex polygons to separate these cells accurately.”

The system tracks two key measurements: positivity index (i.e., proportion of positive cells) and staining intensity index. When tested against expert pathologist assessments using 115 image patches from 26 patients with AML, the AI achieved a 90% correlation with human experts for both indices.

What makes this work particularly practical is how the researchers dealt with limited training data, which is a key challenge in training AI models for pathology. They developed their model using only 35 fully annotated image patches and employed computational techniques to maximize learning from this small dataset.

The system processes slides approximately nine times faster than manual assessment. For busy pathology laboratories, this could mean faster turnaround times without compromising quality. This may also free pathologists to focus on more complex diagnostic tasks.

Looking ahead, Zabihollahy sees broader applications: “We’re hoping to adapt this approach for other biomarkers, like Ki67 in breast cancer. The basic principles should transfer well.”

The team is now planning studies to correlate their AI-based assessments with patient outcomes, which is a crucial step toward clinical validation. They are also working on making the system more robust across different laboratory conditions and staining protocols.Fatemeh Zabihollahy, PhD, lead AI scientist at the University of Toronto, Canada.

[1] Fatemeh Zabihollahy, Automated quantification of TP53 using digital immunohistochemistry for acute myeloid leukemia prognosis. Presented at SPIE 2025 Digital and Computational Pathology conference, February 18, 2025; San Diego, CA.