AI Cancer Tools: Risks of ‘Shortcut Learning’ & Unreliable Diagnoses

The promise of artificial intelligence to accelerate cancer diagnosis and reduce the burden on pathology services is facing a critical re-evaluation. New research published in Nature Biomedical Engineering suggests that many AI systems designed to predict cancer biomarkers from microscope images may be relying on visual “shortcuts” rather than identifying genuine biological signals. This raises concerns about the reliability of these tools for real-world patient care and highlights the need for more rigorous evaluation before widespread adoption.

The study, led by Dr. Fayyaz Minhas of the University of Warwick’s Department of Computer Science, analyzed over 8,000 patient samples across four common cancer types – breast, colorectal, lung, and endometrial. Researchers found that while AI models often achieved high overall accuracy, this performance frequently stemmed from statistical correlations rather than a true understanding of the underlying biology. Essentially, the AI wasn’t learning to identify the cancer itself, but rather recognizing patterns associated with the cancer.

The Restaurant Analogy: Shortcuts vs. Substance

Dr. Minhas illustrates the problem with a relatable analogy: “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.” Many AI pathology models, he explains, are doing the same thing – identifying correlations between biomarkers or obvious tissue features instead of isolating the specific signal of a biomarker. This reliance on shortcuts can lead to inaccurate predictions when conditions change or when biomarkers don’t occur together as expected.

For example, the research team discovered that a model attempting to detect mutations in the BRAF gene might instead learn to identify the presence of microsatellite instability (MSI), another clinical feature often found 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 only accurate when both biomarkers are present, and becomes unreliable when they are not.

Beyond Headline Accuracy: The Importance of Information Gain

Kim Branson, SVP Global Head of Artificial Intelligence and Machine Learning at GSK and a co-author of the study, emphasizes the flawed logic of this approach. “Predicting a BRAF mutation by looking at correlated features like MSI is often like predicting rain by looking at umbrellas—it works, but it doesn’t mean you understand meteorology.”

The researchers argue that a crucial benchmark for evaluating these AI tools is whether they provide information beyond what a pathologist can already discern from routine examination of tissue samples. If a model’s predictive power doesn’t exceed a pathologist’s assessment of tumor grade, it hasn’t truly advanced the field; it has simply automated an existing shortcut.

Their analysis revealed that, for certain prediction tasks, the advantage of deep learning over human assessment was surprisingly modest. AI systems achieved around 80% accuracy in predicting biomarkers, while pathologists using tumor grade alone achieved approximately 75% accuracy. This suggests that, in some cases, the AI is simply replicating existing clinical knowledge, rather than providing novel insights.

Subgroup Analysis Reveals Hidden Dependencies

To further investigate the reliance on shortcuts, the researchers assessed the performance of the AI models within specific patient subgroups. When they focused on only high-grade breast cancers or only MSI-positive tumors, the accuracy of the models significantly decreased. This demonstrated that the models were heavily dependent on the shortcut signals that disappear when confounding factors are controlled.

Professor Nasir Rajpoot, Director of the Tissue Image Analytics (TIA) Center at the University of Warwick and CEO of Histofy, underscores the importance of rigorous evaluation. “This study highlights a critical point about the rollout of AI in medicine: 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.”

Potential Applications and the Path Forward

Despite these limitations, the researchers are not dismissing the potential of machine learning in cancer diagnostics. They acknowledge that these methods can still be valuable for research, drug development candidate screening, and as a tool for clinical triaging, screening, or supplementary decision support. But, they emphasize the need for future AI tools to move beyond correlation-based learning and adopt approaches that explicitly model biological relationships and causal structures. AI-based predictions must be judged through rigorous, bias-aware evaluation.

The study calls for stronger evaluation standards, including subgroup testing and comparison against simple clinical baselines, before these tools are deployed in routine clinical practice. Dr. Minhas concludes, “This research is not a condemnation of AI in pathology. We see 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.”

Co-author, Prof. Sabine Tejpar, Head of Digestive Oncology at KU Leuven, adds a cautionary note about the rush to innovation in oncology. “Clinical relevance of novel tools requires grounded tailoring to what is precise, correct and feasible for the individual patient. Too often, oncology is swept up by ‘innovation’ with limited or no impact on patient care, driven more by what can be provided or sold than by rigorous assessment of what is truly relevant for individual patients and their specific features.”

Next Steps: Stricter Evaluation and Causal Modeling

The future of AI in pathology hinges on developing algorithms that can learn causal relationships rather than simply identifying correlations. This requires a shift in focus from building larger and more complex models to implementing stricter evaluation protocols that force algorithms to demonstrate genuine biological understanding. The research team advocates for a more cautious and evidence-based approach to the implementation of AI in cancer diagnostics, prioritizing patient safety and clinical validity over headline accuracy.