AI Cancer Detection: Promising Tool or Statistical Shortcut?

Artificial intelligence is rapidly transforming healthcare, with AI-powered tools increasingly used to analyze medical images and predict cancer diagnoses. However, new research from the University of Warwick, published in Nature Biomedical Engineering, suggests that many of these systems may be relying on misleading visual cues rather than genuine biological signals, raising concerns about their reliability in real-world clinical settings.

The Problem of “Shortcut Learning”

The study, led by Dr. Fayyaz Minhas, Associate Professor and principal investigator of the Predictive Systems in Biomedicine (PRISM) Lab at the University of Warwick, analyzed over 8,000 patient samples across four major cancer types: breast, colorectal, lung, and endometrial. Researchers found that while AI models often achieve high accuracy rates, this performance frequently stems from identifying statistical “shortcuts” rather than understanding the underlying biology of the disease.

“It’s a bit like judging a restaurant’s quality by the queue of people waiting to get in: it’s a useful shortcut, but it’s not a direct measure of what’s happening in the kitchen,” explains Dr. Minhas. “Many AI pathology models are doing the same thing, relying on correlations between biomarkers or on obvious tissue features, rather than isolating biomarker-specific signals. And when conditions change, these shortcuts often fall apart.”

How Shortcuts Manifest

For example, the research team discovered that a model attempting to detect mutations in the BRAF gene – a gene associated with cancer – might instead learn to identify the presence of microsatellite instability (MSI), a clinical feature often occurring alongside BRAF mutations. The AI then predicts BRAF status based on the presence of MSI, rather than directly detecting the BRAF mutation itself. This means the prediction is accurate only when both biomarkers co-occur, becoming unreliable when they do not.

Kim Branson, SVP Global Head of Artificial Intelligence and Machine Learning at GSK and a co-author of the study, likened this to predicting rain by observing umbrellas. “It works, but it doesn’t mean you understand meteorology,” he stated. He emphasized that AI tools must demonstrate information gain beyond what a pathologist can already assess through traditional methods to truly advance the field.

Accuracy Declines with Subgroup Analysis

When the AI models were tested on specific patient subgroups – such as high-grade breast cancers or MSI-positive tumors – accuracy significantly decreased. This revealed the models’ dependence on shortcut signals that disappear when confounding factors are controlled. In some cases, the performance advantage of deep learning over traditional clinical assessments was minimal. AI systems achieved just over 80% accuracy in predicting biomarkers, compared to around 75% using tumor grade alone – a metric already evaluated by pathologists.

The Path Forward: Rigorous Evaluation and Biological Modeling

Despite these concerns, researchers emphasize that machine learning still holds value in areas like research, drug development, and clinical triaging. However, they argue that future AI tools must move beyond correlation-based learning and focus on explicitly modeling biological relationships and causal structures.

Professor Nasir Rajpoot, Director of the Tissue Image Analytics (TIA) Centre at the University of Warwick and CEO of Warwick spin-out Histofy, highlighted the importance of rigorous, bias-aware evaluation. “To deliver real and lasting impact, the value of AI-based clinically important predictions must be judged through rigorous, bias-aware evaluation, rather than relying solely on headline accuracies that fail to account for confounding effects.”

Dr. Minhas concludes, “This research is not a condemnation of AI in pathology. It is a wake-up call. Current models may perform well in controlled settings but rely on statistical shortcuts rather than genuine biological understanding. Until more robust evaluation standards are in place, these tools should not be seen as replacements for molecular testing, and it is essential that clinicians and researchers understand their limitations and use them with appropriate caution.”

Coauthor, Prof. Sabine Tejpar, Head of Digestive Oncology at KU Leuven, added that clinical relevance requires tailoring tools to individual patient needs and avoiding oversimplification or overreach.

Key Takeaways

• AI cancer detection tools show promise but may rely on misleading visual cues.
• “Shortcut learning” can lead to inaccurate predictions when conditions change.
• Rigorous evaluation and biological modeling are crucial for developing reliable AI tools.
• AI should not replace traditional molecular testing but can serve as a supplementary tool.

Source: Dawood, M., et al. (2026). Confounding factors and biases abound when predicting molecular biomarkers from histological images. Nature Biomedical Engineering. DOI: 10.1038/s41551-026-01616-8