Many diseases remain difficult to treat because their earliest signals are weak, scattered, or invisible to conventional tests. For clinicians, the challenge is not only to detect disease earlier, but also to see it more clearly in complex medical images, and to act faster when decisions must be made during treatment or surgery. Ovarian cancer, for example, is often diagnosed only after it has advanced; pathology AI still depends heavily on expert-labeled images; and intraoperative cancer diagnosis can still require frozen-section workflows that take about 30 minutes.

This challenge is motivating Prof. Zhang Kang’s research team at Macau University of Science and Technology to develop AI-driven approaches that can uncover weak biological signals, localize disease more clearly in medical images, and accelerate diagnosis at critical clinical moments.

Across cancer detection, pathology imaging, surgical diagnosis, biological aging, reproductive health, vision disorders, and treatment prediction, the team is building what can be described as AI-driven early-warning systems for medicine — tools that help clinicians see disease earlier, faster, and more clearly.

One of the team’s most clinically significant advances focuses on seeing cancer earlier, before it becomes easier to detect by conventional methods. In a study published in Cell Reports Medicine¹, the researchers analyzed methylation patterns in cell-free DNA — tiny DNA fragments circulating in the bloodstream — to identify ovarian cancer-associated changes. The study screened more than 3.3 million methylation sites and validated the approach in 754 ovarian cancer patients, including 205 early-stage cases, and 1,118 healthy controls. By learning how weak methylation signals relate to one another, the system achieved around 80 percent sensitivity and 95 percent specificity for early-stage ovarian cancer detection. The researchers also translated key markers into a droplet digital PCR assay, suggesting a possible pathway toward more practical blood-based screening for one of the deadliest gynecological cancers.

Seeing disease earlier also depends on seeing it more clearly once medical images are available. In a 2026 study published in Nature Biomedical Engineering², the team developed a multimodal vision–language model that can localize disease-relevant regions in medical images without requiring extensive manual labeling by specialists. The model learns from the relationship between medical images and clinical reports, allowing it to identify suspicious regions across chest X-rays, retinal images, and histopathology slides. This could make pathology and radiology AI more scalable, especially in settings where expert-labeled data are limited.

The third challenge is speed at the point of care. In breast cancer surgery, clinicians often rely on frozen-section pathology to assess suspicious tissue, a process that can take roughly 30 minutes and interrupt surgical workflow. In a study published in Science Bulletin³, the team combined dynamic optical imaging with learning-based analysis to distinguish malignant from non-malignant breast tissue in approximately three minutes. In prospective breast cancer cohorts, the approach achieved diagnostic accuracy above 97 percent. Rather than depending on chemical staining or labor-intensive tissue preparation, the system analyzes microscopic structural patterns directly from freshly excised tissue, potentially enabling faster surgical decisions and reducing waiting time in the operating room.

Beyond these cancer-focused studies, the team is extending the same logic across increasingly diverse medical materials — from pathology slides and retinal images to electronic health records and molecular data. Other projects integrate imaging, clinical records, and biological information to uncover hidden subtypes of breast cancer, predict responses to immunotherapy, and estimate long-term outcomes from non-invasive scans. According to the team, the broader goal is not simply automation, but helping clinicians connect weak or scattered signals that are difficult to interpret in isolation.

Reference:

1) https://www.cell.com/cell-reports-medicine/fulltext/S2666-3791(24)00380-X

2) https://www.nature.com/articles/s41551-025-01574-7

3) https://www.sciencedirect.com/science/article/pii/S2095927324002172?via%3Dihub

(Newswise/HG)