A new AI system can predict lymphoma subtypes and order diagnostic tests directly from initial biopsy scans, cutting down the days patients spend waiting for a diagnosis.

Why do cancer patients wait days just to get their diagnostic tests ordered? The delay is rarely the lab work itself. Instead, the bottleneck lies in the manual back-and-forth of pathology workflows. A specialist must preview an initial tissue slice, guess the likely lymphoma subtype, and then order specific chemical stains to confirm it.

This trial of a new AI tool called HATS suggests we can automate this middle step entirely. By predicting the right tests instantly, the system challenges the assumption that human clinical judgment must drive every step in the lab. It suggests that pattern recognition can safely take over the initial logistics of cancer diagnostics.

How the system works

Researchers built the Hematopathology Automatic Triaging System, or HATS, by evaluating seven public pathology foundation models. They trained the system on 4,996 whole-slide images from 1,607 patients. This dataset covered the 10 most common lymphoma categories, making it a highly realistic test of clinical utility.

In a blinded study, the AI proved remarkably sharp. When forced to guess lymphoma subtypes from tissue structure alone, HATS achieved 85% accuracy. Practicing pathologists scored just 65% on the same task. This gap shows that machine learning can spot subtle cellular patterns that human eyes simply miss.

The real-world test

Predicting subtypes on paper is one thing, but ordering the right tests in a busy clinic is another. The study evaluated how well the AI translated its predictions into actual medical orders. Here is how the numbers stacked up:

• The system reached 84% case-level subtype classification accuracy, with a 0.962 ROC-AUC.
• This performance translated to a 92% accuracy rate in ordering the correct immunohistochemistry panels.
• In an independent real-world validation of 230 clinical cases, the AI-ordered panels were sufficient for a final diagnosis in 72.6% of patients.

The system did not work perfectly on every sample. In the real-world test, it flagged 7 cases with scant tissue and safely routed them to human pathologists for manual review. This safety valve is crucial for clinical safety.

The clinical reality

Skeptics will point to the 72.6% sufficiency rate in the real-world test. If nearly thirty percent of patients still need extra tests, does this actually save time? The answer is yes, because of how clinical workflows operate.

In nearly three-quarters of cases, the AI gets the diagnostic panel right on day one. This bypasses a major administrative bottleneck. For the remaining cases, pathologists still review and adjust the orders. No patient is left with an incomplete diagnosis, but the majority get their results days faster.

This is where clinical AI is heading. Instead of trying to replace the doctor, software is taking over the tedious, high-volume decisions that delay patient care. By automating the triaging step, labs can speed up diagnoses without losing human oversight.

Read the full study in medRxiv.