New approach may improve AI detection of cancer in tissue samples

Many AI systems used in pathology divide tissue slides into fixed sections and analyze them separately. Pathologists, by contrast, use a more flexible approach, initially scanning the entire slide before changing magnification and moving between regions, then focusing on areas that appear abnormal.

This distinction is important because a tissue slide can contain billions of pixels, while cancer-related abnormalities may be confined to a very small area.

Chi Huang, an assistant professor of pathology and laboratory medicine at the University of Pennsylvania and a co-author of the study, compared the process to a helicopter searching for missing people. Rather than examining a small area immediately, a pathologist first surveys the entire scene before conducting a detailed search.

Huang and his colleagues developed a method to train AI to replicate the search behavior of pathologists rather than learning solely from their final diagnostic conclusions.

AI learns from pathologists’ search patterns

The researchers collected data from eight pathologists, tracking their movements across tissue slides and the magnification levels they used while searching for suspicious regions. They then filtered out incidental movements and focused on behaviors indicating genuine attention to a particular area, such as spending more time there or examining it repeatedly.

The researchers compared these data with eye-tracking measurements to verify that the system was learning the regions on which pathologists actually focused. They also asked the AI model to provide a brief explanation of why each region was considered important, allowing pathologists to accept, modify or reject the explanation.

Using these data, the team developed a system called Pathology-o3. The system initially scans a slide at low resolution, identifies regions that warrant closer examination, and then sends high-resolution images of those areas to an AI model for analysis.

Promising results, but limitations remain

The researchers tested Pathology-o3 on slides containing lymph-node tissue from patients with colorectal cancer and compared its performance with OpenAI’s o3 model.

Pathology-o3 identified all slides containing cancer, achieving 100% sensitivity for positive cases. However, 15.5% of the slides it classified as positive were actually cancer-free.

The o3 model identified cancer-positive slides with 87.5% accuracy, but 53.3% of the slides it classified as positive were actually negative.

When tested on an independent dataset it had not previously encountered, Pathology-o3 achieved 97.6% accuracy in identifying cancer-positive slides, while 37.1% of cases it classified as positive were actually negative.

The researchers suggest that the relatively high rate of false positives may be linked to the system’s design, which favors selecting regions for further examination in order to minimize the risk of missing cancerous features.

Mohammad Asadi, a data scientist at Stanford University who was not involved in the study, said the findings suggest the system could potentially be applied to slides from different sources. However, he noted that the results do not yet demonstrate that pathologists would become more accurate or faster when using the system.

Could the system assist pathologists?

The study was not designed to determine whether Pathology-o3 could outperform pathologists or diagnose cancer independently. Instead, the researchers sought to assess whether training AI to replicate the way pathologists search for abnormalities could improve its performance.

Huang said the main value of the approach lies in making use of data that hospitals already collect but that have not been sufficiently incorporated into AI training.

The researchers plan further trials to determine whether using the system alongside pathologists could help them detect more cancer cases and work more efficiently.

Asadi said the most practical application at this stage would be to use the system as a preliminary screening tool that directs pathologists toward areas requiring closer examination. However, he stressed the need for multicenter trials to assess the system’s accuracy and speed, as well as the number of false alarms and the additional workload they could create for clinicians.

The system also remains limited compared with the full diagnostic process, which may require analysis of multiple tissue slides, different staining techniques and the patient’s medical history.

For this reason, Huang emphasized that the goal is not for AI to take over diagnosis independently, saying: “I would not claim that it should make the diagnosis on its own.”