A new study reveals that training artificial intelligence to mimic the way pathologists search for signs of cancer may enhance its ability to detect suspicious areas within tissue samples.
اضافة اعلان
Many AI systems used in pathology rely on dividing a tissue slide into fixed sections and analyzing them, while doctors follow a more flexible approach: they begin by scanning the slide broadly, then change the zoom level and move between different areas, before focusing on parts that appear abnormal.
The importance of this method lies in the fact that a tissue slide may contain billions of pixels, while the signs indicating the presence of cancer may be confined to a very small area.
Chi Huang, assistant professor of pathology and laboratory medicine at the University of Pennsylvania and one of the study's co-authors, compared this process to a helicopter searching for missing people, explaining that a doctor doesn't start by examining a small patch of ground, but scans the whole scene before moving to close inspection.
Huang and colleagues developed a method to train AI on the behavior of pathologists while examining slides, rather than training it only on the final results they arrive at.
AI Learns from the Doctor's Method
The researchers gathered data from eight specialists, recording their movements across the slides and the zoom levels they used while searching for suspicious areas. They then excluded incidental movements and focused on behaviors reflecting genuine interest in a particular area, such as pausing longer or repeatedly examining a region.
The researchers compared this data with eye-tracking data, to confirm that the system was learning the areas doctors were actually focusing on. They also asked the AI model to provide a brief explanation of why each area was significant, giving doctors the opportunity to accept, modify, or reject the explanation.
The researchers used this data to build a system they named "Pathology-o3," which begins by scanning the slide at low resolution, then identifies areas warranting closer examination, and sends high-resolution images of these areas to an AI model for analysis.
Promising Results, But Not Yet Sufficient
The researchers tested "Pathology-o3" on slides containing lymph node tissue from patients with colorectal cancer, comparing its performance to OpenAI's "o3" model.
"Pathology-o3" was able to identify all the slides containing cancer, achieving 100% accuracy in detecting positive cases. However, 15.5% of the slides it classified as positive were actually cancer-free.
The "o3" model, meanwhile, achieved 87.5% accuracy in identifying affected slides, but 53.3% of the slides it classified as positive were actually negative.
When the system was tested on an independent dataset it had not previously encountered, its accuracy in identifying cancer-affected slides reached 97.6%, while 37.1% of the cases it classified as positive were actually negative.
The researchers believe the high rate of false positives may be related to the system's design, which tends to select areas warranting additional examination in order to reduce the likelihood of missing signs of cancer.
Mohammad Asadi, a data scientist at Stanford University who was not involved in the study, said these results suggest the system could potentially be used with slides from different sources, but they do not yet prove that doctors will become more accurate or faster when using it.
Can It Help Doctors?
The study did not aim to prove that "Pathology-o3" can outperform pathologists or diagnose cancer on its own, but rather to test whether training AI on a doctor's search method could improve its performance.
Huang believes the core value of this approach lies in making use of data that already existed within hospitals but had not been sufficiently used in training AI.
The researchers plan to conduct follow-up trials to determine whether using the system alongside pathologists could help them detect a greater number of cancer cases and work more quickly.
Asadi believes the most sensible use of the system at present is as an initial screening tool, directing the doctor to areas warranting further scrutiny. However, he stresses the need for trials involving multiple hospitals, to measure the system's accuracy and speed, as well as the rate of false alarms and the workload it might add for doctors.
The system also remains limited compared to the full diagnostic process, which may require analyzing several slides and using different staining techniques, in addition to reviewing the patient's medical history. For this reason, Huang emphasizes that the goal is not for AI to handle diagnosis on its own, saying: "I would not claim that it should make the diagnosis by itself."
Resource: Al-Ghad.
