Cancer diagnoses directly affect treatment decisions and prognosis and can even have life-altering consequences. Traditional pathological diagnosis primarily relies on pathologists examining and analysing tissue slides under a microscope and making judgements based on their professional experience. In recent years, the application of artificial intelligence (AI) in pathology diagnosis (pathology AI) has advanced rapidly, helping to improve diagnostic efficiency and support clinical decision-making. However, many existing AI models still lack a complete and verifiable mechanism to ensure the reliability of diagnostic outcomes, limiting the application of such technology in high-stakes clinical settings.
To tackle this challenge, Prof. ZHANG Xiaoge, Assistant Professor of the Department of Industrial and Systems Engineering at The Hong Kong Polytechnic University (PolyU), and his research team have developed an integrated AI framework named TRUECAM (TRustworthiness-focused, Uncertainty-aware, End-to-end CAncer diagnosis with Model-agnostic capabilities). Applied to whole-slide image analysis for non-small cell lung cancer subtyping, the framework ensures both data and model trustworthiness, laying an important foundation for the safe application of pathology AI in cancer diagnosis.
TRUECAM is designed to enhance the reliability of AI-assisted cancer diagnosis. It can assess the level of the AI's confidence in its diagnostic outputs and proactively prompt pathologists to review cases when uncertainty is high or when input data fall outside the model's scope. At the same time, the framework is model-agnostic, supporting the complete analysis pipeline from pathology images to diagnostic outcomes and helping healthcare professionals apply AI-generated diagnoses more reliably.
The framework is applied to whole-slide imaging, a process that digitally scans glass tissue slides into high-resolution digital images and provides "virtual microscopy", allowing pathologists to review pathology samples on a computer. The findings show that TRUECAM is applicable not only to non-small cell lung cancer subtyping, but can also be extended to breast, brain, and kidney cancer subtyping tasks, as well as a 46-class pan-cancer slide-level classification setting, demonstrating its broad application potential and clinical translational value.
As a general framework, TRUECAM can be integrated into pathology AI models of various sizes, architectures, purposes and complexities to support responsible clinical applications. The framework has three core functions: detecting out-of-scope inputs, automatically eliminating highly ambiguous and difficult-to-judge image regions, and applying conformal prediction to keep diagnostic error rates within an acceptable range.
The research team conducted a systematic evaluation of the framework across multiple cancer datasets using two types of AI models: specialised models designed for particular tasks and foundation models with broad application potential. Their computational experiments indicate that TRUECAM-wrapped models consistently outperformed their unwrapped counterparts in classification accuracy, robustness, interpretability, data efficiency and fairness.
Prof. Zhang said, "TRUECAM strikes a sound balance between fully pathology AI-powered and purely pathologists-led cancer diagnosis. When the model's confidence in its diagnostic outputs is high, the system can help handle clear-cut cases, while uncertain cases are flagged and passed on to pathologists for further review and clinical judgement. This AI–pathologist collaboration helps improve diagnostic efficiency, lighten pathologists' workloads, enhance diagnostic reliability and scale up diagnostic capacity."
Prof. Zhang added, "While the published study focuses mainly on whole-slide images, we are further exploring additional modalities, such as molecular profiles including RNA-sequencing and diagnostic reports, to broaden the framework's potential scope. TRUECAM provides a systematic solution to building trustworthy pathology AI and strengthens the foundation for deploying it in real-world settings."
This research has been published in the internationally renowned journal Nature Biomedical Engineering, and received funding support from the National Natural Science Foundation of China, the Research Grants Council of the Hong Kong Special Administrative Region, and the Shenzhen Science and Technology Program.
