AI Collaboration Enhances Cancer Diagnosis Accuracy and Reduces Pathologist Workload
Artificial intelligence is poised to revolutionize cancer diagnosis by learning to defer to human pathologists strategically, thereby improving diagnostic accuracy and alleviating the burden on overstretched medical professionals. New research from the University of Surrey and Monash University demonstrates a probabilistic method that allows AI systems to learn from incomplete expert input whereas ensuring equitable workload distribution among teams.
Addressing Critical Challenges in AI-Assisted Diagnosis
Current human-AI collaborative systems often require exhaustive review of every case by experts during training, a process that is both costly and time-consuming. These systems can inadvertently overload the most skilled experts during testing, potentially leading to burnout and diagnostic errors. The new approach tackles these limitations head-on.
How the New System Works
The research introduces a probabilistic method enabling AI systems to learn effectively even with incomplete expert input. This system distributes workload evenly across teams, preventing the overwork of individual pathologists. Testing on colon cancer pathology images revealed that the system maintained high accuracy even when 70% of expert annotations were missing, all while preventing any single pathologist from being overwhelmed.
“In cancer pathology and radiology, we know that overloading experts leads to mistakes. There is a documented case where a radiologist misdiagnosed as they interpreted 162 cases in one day when the average is only 50,” explains Professor Gustavo Carneiro, co-author of the study from the Centre for Vision, Speech and Signal Processing at the University of Surrey. “Our system prevents this by ensuring work is distributed fairly while maintaining high accuracy. The AI learns to handle routine cases independently and defer complex ones to humans, but crucially, it doesn’t always defer to the same person.”
Beyond Colon Cancer: Versatility Across Medical Imaging
The challenge of increasing caseloads is particularly acute in cancer diagnosis, where differentiating between benign, precancerous, and malignant tissue demands specialized expertise. An AI system capable of confidently managing straightforward cases and flagging complex ones for human review can significantly reduce pressure on specialists without compromising diagnostic precision. The research team also successfully tested the approach on chest X-ray interpretation and bone disease imaging, highlighting its adaptability across various medical imaging tasks.
Dr. Cuong Nguyen, lead author and researcher at Surrey’s Centre for Vision, Speech and Signal Processing, notes, “Previous systems assumed you could gain every expert to review every training sample, which simply is not realistic for large datasets or busy clinical teams. We have shown you can train effective Human-AI systems even when experts only review portions of the data. This makes the technology far more practical for real-world deployment in cancer pathology and other high-stakes medical fields.”
Algorithm and Workload Management
The system employs an algorithm that considers both the selection of which expert to consult and any missing expert opinions as variables to be inferred during training. It also incorporates a mechanism to regulate workload assignment to each expert and the AI classifier itself, allowing organizations to establish workload limits during training rather than making adjustments afterward.
Addressing Concerns About AI in Healthcare
This research addresses growing concerns surrounding AI implementation in healthcare, where fully automated systems may overlook crucial details, while relying on human consultation for every decision is impractical and expensive. The system represents a balanced approach, enhancing diagnostic capabilities without replacing the essential role of human expertise.
The research was presented at the International Conference on Learning Representations (ICLR) 2025.