A recent study indicates that artificial intelligence (AI) algorithms for cancer detection could perform more effectively by mimicking the dynamic search methods of human pathologists rather than analyzing predetermined regions of tissue samples. Researchers from the University of Pennsylvania have developed a new training approach, termed “Pathology-CoT” (Chain of Thought), which teaches AI how pathologists move and focus on areas of interest in pathology slides.
While traditional AI systems often train on static images labeled for cancer presence, Pathology-CoT leverages the actual search behaviors of pathologists, recording movements and adjustments in magnification to gather more relevant training data. The study revealed that an AI tool called Pathology-o3 could correctly identify cancerous slides 100% of the time but had a higher false-positive rate compared to other general AI systems.
The researchers emphasize that this AI tool is not intended for independent diagnoses but rather as a prescreening aid, directing pathologists’ attention to specific areas of concern. Further research is planned to determine how the use of Pathology-o3 affects pathologists’ accuracy and efficiency in diagnosing cancer.
