AI Tumor Classifier Proves Its Worth in 516 Tough Brain Cancer Cases

Every year, thousands of patients with brain tumors receive diagnoses that determine which therapies they will undergo, how their surgeons will proceed, and what their prognosis is likely to be. Increasingly, those diagnoses rest not only on what a pathologist sees under the microscope, but on the epigenetic fingerprint of the tumor itself. DNA methylation profiling, which charts chemical tags scattered across the genome, has matured into one of the most powerful classification tools in neuropathology, capable of distinguishing tumor entities that look nearly identical histologically yet behave in profoundly different ways. Now, a new study published in Acta Neuropathologica has put one of the leading machine learning classifiers through one of its most demanding real-world tests to date, examining how it performs across 516 central nervous system tumor specimens ranging from the routine to the diagnostically hopeless.

The study, led by Charlotte Brandenburg and Leonille Schweizer of the Edinger Institute at University Hospital Frankfurt, together with colleagues including Kenneth D. Aldape and Omkar Singh of the National Cancer Institute in Bethesda, Maryland, evaluated the Bethesda Classifier version 3, or Bv3, a random forest-based machine learning tool trained on DNA methylation data from CNS tumors. The central question was deceptively simple: how well does Bv3 perform not just on clean, textbook cases, but on the murky, ambiguous specimens that haunt every neuropathology laboratory? The answer, the researchers report, is that Bv3 performs robustly in both settings, and that its greatest value may lie precisely in the cases where other classifiers throw up their hands.

The team structured their evaluation in two cohorts. The first consisted of 195 CNS tumor samples that had already been successfully classified by the Heidelberg Classifier v12.8, or Hv12.8, one of the most widely used methylation-based classifiers in the field and the reference standard in many European centers. These represented the straightforward cases, the samples where the methylation pattern aligned cleanly with a known tumor class and the calibrated score left little room for doubt. The second cohort was far more interesting from a clinical standpoint: 321 retrospectively collected cases that Hv12.8 had failed to classify at all, the diagnostically challenging specimens that had remained unclassifiable despite integrated histological and molecular workup.

To judge performance, the researchers first re-evaluated both cohorts to establish an integrated histomolecular diagnosis according to the WHO Classification of Tumours of the Central Nervous System, fifth edition, published in 2021, which they refer to as the final WHO2021 diagnosis. Misclassification was defined conservatively: a case counted as misclassified only when Bv3 assigned a class with a calibrated score of at least 0.9, indicating high confidence, and that assignment did not match the final WHO2021 diagnosis. This threshold matters because methylation classifiers report a probability-like score, and only high-scoring outputs are typically acted upon in clinical practice. A confident wrong answer is far more dangerous than a hesitant or absent one.

In the straightforward cohort, the results were strikingly reassuring. Bv3 correctly classified 189 of the 195 cases, a concordance rate of 97 percent, essentially confirming that the tool performs at the level of the established Heidelberg classifier when cases are unambiguous. For laboratories considering whether to invest in a second, independent methylation classifier, this figure addresses the most basic concern: Bv3 does not introduce errors where existing tools already succeed. The two classifiers, built by different groups on overlapping but distinct training sets, agree almost perfectly on the easy cases, which is exactly what one would hope to see from a validated diagnostic instrument.

The challenging cohort told the more compelling story. Of the 321 cases that Hv12.8 could not classify, Bv3 managed to assign a confident class to 180, or 56 percent. In other words, the Bethesda classifier rescued more than half of the specimens that one of the field’s flagship tools had left in diagnostic limbo. This is not a trivial gain. Unclassifiable CNS tumors are a genuine clinical problem; studies have shown that patients whose tumors cannot be assigned to a methylation class face diagnostic uncertainty, and in some contexts worse prognostic counseling, because integrated diagnosis according to WHO CNS5 often depends on methylation class assignment. A tool that converts half of those dead ends into actionable diagnoses changes the calculus of routine molecular neuropathology.

Accuracy, of course, is only as valuable as it is reliable, and here the study offers both reassurance and caveats. Among the 180 newly classified challenging cases, misclassification with high confidence occurred in 13 of 321 cases overall, roughly 4 percent of the challenging cohort. Notably, more than half of those confident errors, 7 of 13, involved tumors labeled as ganglioglioma by Bv3, a class that appears to act as a magnet for certain ambiguous methylation patterns. The authors also flagged a specific vulnerability: Bv3 assigned the class of high-grade astrocytoma with piloid features, or HGAP, with scores of at least 0.9 in three cases that ultimately did not match the final WHO2021 diagnosis. HGAP, a relatively recently delineated entity combining piloid morphology with IDH-wildtype genetics and MAPK pathway alterations, remains one of the trickier classes for methylation-based assignment, and these findings suggest that high-confidence HGAP calls from Bv3 should be interpreted with particular care and always correlated with histology and molecular genetics.

Glioblastoma, the most common and most lethal primary brain tumor in adults, received its own analysis, and the results carry practical significance for any laboratory handling less-than-perfect tissue. Bv3 correctly identified 79 percent of glioblastomas in the evaluation, and crucially, this success rate was maintained even in tumors with lower DNA concentration and lower tumor purity, the degraded, necrosis-laden, and cell-sparse specimens that are common in surgical neuropathology. Methylation profiling depends on sufficient input DNA and adequate tumor content, and prior work has shown that low input DNA can undermine the reliability of methylation measurements on the Illumina MethylationEPIC platform. The finding that Bv3 retains substantial diagnostic power on imperfect samples suggests a degree of resilience that matters in the real world, where the freshest, most cellular tissue is not always what reaches the profiling pipeline.

The broader message of the study is not that one classifier should replace another, but that complementary deployment is the smart play. Bv3 and Hv12.8 are built on different training cohorts and architectural choices, and their failure modes are, at least in part, non-overlapping. A case that stumps one classifier may be confidently and correctly resolved by the other, and vice versa. The Frankfurt team argues that Bv3 represents a valuable complementary machine learning tool in routine diagnostics, and the numbers support that framing: near-perfect concordance on straightforward cases, a majority rescue rate on unclassifiable ones, and a low but non-zero misclassification rate concentrated in identifiable classes such as ganglioglioma and HGAP. For laboratories running methylation profiling, adding a second calibrated classifier is a relatively low-cost way to raise the diagnostic yield of an already expensive and technically demanding workflow.

The study also arrives at a moment when the field is expanding rapidly. Recent work has pushed methylation classification toward larger training sets, expanded class catalogs, and explainable artificial intelligence approaches that illuminate why a classifier makes a particular call. The WHO CNS5 framework has made histomolecular integration the norm rather than the exception, and methylation profiling has moved from a research curiosity to an expected component of modern CNS tumor diagnostics. Against that backdrop, independent validation of competing classifiers in mixed-difficulty cohorts fills a genuine gap, because published classifier performance is often reported on curated research sets rather than the messy spectrum of routine specimens. This evaluation of 516 cases, spanning both the easy and the seemingly impossible, offers the kind of head-to-head, real-world evidence that laboratories and clinicians need as they decide how to build resilient molecular diagnostic pipelines. It is a reminder that in the age of AI-assisted pathology, the question is no longer whether machine learning belongs in the neuropathology lab, but how best to deploy multiple tools so that each covers the blind spots of the others.

Subject of Research: Evaluation of the Bethesda v3 DNA methylation-based classifier for CNS tumor classification in routine neuropathological diagnostics

Article Title: Complementary use of the DNA methylation-based tumor classification by the Bethesda v3 classifier in routine diagnostics: evaluation in a series of 516 cases

Article References: Complementary use of the DNA methylation-based tumor classification by the Bethesda v3 classifier in routine diagnostics: evaluation in a series of 516 cases. (n.d.). https://doi.org/10.1007/s00401-026-03079-2

Image Credits: AI Generated

DOI: 10.1007/s00401-026-03079-2

Keywords: DNA methylation, brain tumor classification, Bethesda classifier v3, machine learning, neuropathology, CNS tumors, glioblastoma, epigenetics, WHO CNS5 classification, Heidelberg classifier, diagnostic accuracy, tumor profiling