With breast cancer screening demand rising and a global shortage of radiologists, healthcare systems are seeking more efficient ways to interpret mammograms. Artificial intelligence offers a potential solution, either by replacing or supporting human experts. However, the decision to automate must consider both economic and clinical outcomes.
A recent study explored three strategies for mammography interpretation—relying solely on radiologists, using AI alone or combining both through task delegation. Using an optimisation model and data from a large AI challenge, the study examined how disease prevalence, litigation risk and algorithm performance influence which strategy offers the best value for healthcare providers.
Evaluating AI, Radiologists and Delegation
The study compared three approaches: expert-alone, where radiologists assess all mammograms; automation, where AI handles all cases; and delegation, where AI evaluates low-risk cases while high-risk ones are referred to radiologists. The delegation model introduces a decision threshold: cases falling below the threshold are classified as healthy by AI, while those above are reviewed by a human expert. This approach reduces radiologists’ workload and allows them to concentrate on complex assessments.
To determine which strategy is optimal, the authors developed a cost model incorporating expenses for AI use, expert involvement, follow-up procedures and litigation from missed diagnoses. Their findings show that delegation often results in the lowest overall costs. Cost savings ranged from 17.5 percent to 30.1 percent compared with expert-alone solutions, especially when AI performance was moderate to high. Full automation becomes preferable only when AI achieves a very high predictive accuracy. Conversely, expert-alone strategies remain justified when radiologists perform well and litigation risks are high.
Must Read: Patient Trust in AI for Mammogram Screening
The study highlights that the economic trade-off between false positives and false negatives significantly shapes the strategy choice. False positives, which lead to unnecessary follow-up procedures, incur clinical and financial burdens. False negatives, on the other hand, can result in delayed cancer diagnosis and legal claims. This makes accurate classification essential, and the choice of strategy must weigh cost against risk.
The Role of Prevalence and Regulation
A key insight from the study is the role of disease prevalence. In populations where breast cancer is more common, the cost implications of missed diagnoses grow, often shifting the balance toward more human involvement. In low-prevalence settings, where fewer positive cases are expected, the relative benefit of automation increases. The study introduces a threshold prevalence ratio, determined by the relative costs of false positives and false negatives. This ratio helps identify when disease prevalence is high enough to warrant a shift in strategy.
Legal and regulatory factors also play a critical part. The study considers scenarios where litigation costs differ for human versus AI decisions. If AI is held to stricter liability standards—as might be the case under product liability laws—then healthcare providers may favour strategies that involve more human oversight. In this context, delegation becomes a compromise, balancing the efficiency of automation with the perceived safety of human judgement.
Policy decisions about liability, therefore, could influence how quickly AI is integrated into clinical workflows. Stricter standards for AI may slow adoption or reinforce the need for hybrid models, even when AI performs well. The authors suggest that such policies could be used to manage the pace of automation in sensitive areas like cancer screening, where decisions have high stakes.
Real-World Validation and Implementation
To validate their model, the researchers used data from the Digital Mammography DREAM Challenge, which involved over 25,000 mammograms assessed by AI algorithms and radiologists. The results confirmed the theoretical predictions. Delegation was consistently the most cost-effective strategy, especially when using top-performing algorithms. Even under scenarios with high AI and litigation costs, the best AI tools still delivered meaningful savings when used in a delegation model.
The practical advantage of delegation lies in its flexibility. Healthcare providers can adjust the threshold at which AI refers a case to a radiologist, tailoring it to their specific population, cost structure and regulatory context. This adaptability makes delegation a scalable solution for systems dealing with increased screening volumes and constrained specialist availability.
Moreover, the delegation approach supports gradual integration of AI into medical practice. Rather than replacing radiologists outright, it allows them to remain central to the diagnostic process, focusing their expertise where it matters most. This not only ensures patient safety but also builds trust in AI tools among clinicians and patients.
The study offers a clear framework for understanding how healthcare organisations can use AI to reduce the cost of mammography screening while maintaining clinical quality. Rather than advocating for full automation, the findings support a shared approach where AI handles routine cases and radiologists address more complex ones. This task-sharing strategy, or delegation, proves to be the most cost-efficient in most scenarios, particularly where AI performs reliably and disease prevalence is moderate.
Source: nature communications
Image Credit: iStock
