Technology
A recent study reveals an innovative approach to training artificial intelligence systems to mimic pathologists' methods in examining tissue samples, potentially significantly enhancing AI's ability to detect suspicious cancer indicators. This research highlights current diagnostic challenges, offering a solution for more accurate and efficient tools. Most AI systems currently used in pathology rely on dividing a tissue slide into fixed, uniform segments for analysis. In contrast, pathologists employ a more flexible approach; they typically begin with a general scan of the entire slide, then adjust magnification levels and navigate between different regions before focusing on areas that appear abnormal. The importance of this human approach stems from the fact that a tissue slide can contain billions of pixels, while the subtle signs of cancer might be confined to a very tiny area. Professor Chi Huang, an Assistant Professor of Pathology and Laboratory Medicine at the University of Pennsylvania and a co-author of the study, likened this process to a helicopter searching for missing persons. Huang explained that, similar to a helicopter pilot, a pathologist does not start by examining a small, specific patch of ground directly but rather scans the entire scene first to get an overview before proceeding to a detailed examination of potential areas. To develop this approach, researchers collected meticulous data from eight pathology specialists, recording their movements within the slides and the magnification levels they used when searching for suspicious regions. The research team then filtered out accidental movements, focusing on behaviors that reflected genuine interest in a particular area, such as pausing for a longer period or repeatedly examining a specific region. Researchers compared this data with eye-tracking information to confirm the system was learning the areas on which the doctors focused. The AI model was also asked to provide a brief explanation for each identified area's importance, allowing pathologists the opportunity to accept, modify, or reject it. Using this data, researchers built a new AI system named "Pathology-o3." This system operates by initially scanning the tissue slide at a low resolution, then identifying areas that warrant more detailed examination. High-resolution images of these identified areas are then sent to another specialized AI model for in-depth analysis. **Promising Results, Yet Requiring Further Evaluation** Researchers tested "Pathology-o3" on slides containing lymph node tissues from patients with colorectal cancer, comparing its performance against "o3" from "OpenAI." "Pathology-o3" demonstrated a remarkable ability to identify all tissue slides containing cancer, achieving 100% accuracy in detecting positive cases. However, 15.5% of the slides it classified as positive were actually cancer-free. In contrast, the "o3" model achieved 87.5% accuracy in identifying infected slides but registered a significantly higher rate of false positives, with 53.3% of the slides it classified as positive being negative in reality. When tested on an independent dataset it had not previously encountered, "Pathology-o3"'s accuracy in identifying cancer-infected slides reached 97.6%. In this dataset, 37.1% of the cases it classified as positive were actually negative. Researchers suggest that this higher rate of false positives might be linked to the system's design, which tends to select more areas for additional examination to minimize the chance of missing any potential signs of cancer. Commenting on these findings, data scientist Mohammed Asadi from Stanford University, who was not involved in the study, noted that these results indicate the promising potential for using the system with slides from diverse sources. However, he emphasized that these findings do not yet prove that pathologists will become more accurate or faster when using this system in their daily practices. **Future Prospects for Medical Assistance** The study's primary objective was not to demonstrate that "Pathology-o3" could outperform human pathologists or diagnose cancer independently. Instead, it aimed to test whether training AI to mimic a physician's search method could improve its performance in identifying suspicious areas. Huang believes that the core value of this approach lies in its ability to leverage vast amounts of medical data already existing within hospitals, which has not yet been adequately utilized in training AI systems. Researchers plan to conduct subsequent experiments to determine whether using this system alongside human pathologists can help them detect a greater number of cancer cases and work at a faster pace. Asadi believes the most logical current use for the system is as a preliminary screening tool, guiding the pathologist directly to areas requiring further scrutiny. Nevertheless, he stresses the critical need for broader clinical trials involving multiple hospitals to measure the system's actual accuracy and speed, as well as to evaluate the volume of false alarms and the potential additional workload it might impose on pathologists. Furthermore, the system remains limited compared to the complete cancer diagnosis process, which often requires analysis of multiple slides, various staining techniques, and a thorough review of the patient's medical history. For this reason, Huang repeatedly affirms that the goal is not for AI to undertake the diagnostic process alone, stating clearly: "I would not claim that it should make diagnoses on its own."
