Before a tumor surgery begins, what a doctor needs most is to know as much as possible about what lies ahead. Is the mass behind the sternum merely adjacent to the pericardium and blood vessels, or has it already grown into them? Can it be completely resected? Does the patient need other treatment first? These questions shape the surgical plan, yet reliable answers often only come from postoperative pathological examination. According to Zhitong Finance APP, the thymus research that DIAGENS-B (02526) is participating in is attempting to bring valuable information forward to the preoperative stage. This is also the most direct entry point for understanding AI for Science: letting AI help researchers process information that is difficult to analyze one by one manually, testing the links between imaging and disease, and making more medical questions affordable and faster to investigate.

Specifically, in this study, the Masaoka-Koga staging in the paper can be understood as a "tumor boundary-crossing map": Stage I remains within the capsule; Stage II begins to break through the boundary, with some invasion requiring microscopic confirmation; Stage III invades adjacent structures such as the pericardium, lungs, or major blood vessels; Stage IV involves dissemination or metastasis. This map affects the difficulty of complete resection, surgical preparation, and treatment sequencing. The challenge is that on CT, "being very close" does not equal tissue already being invaded. Final staging usually depends on surgical and pathological information. Patients who can undergo complete resection and whose physical condition allows it are usually prioritized for surgery; if initial complete resection is difficult to achieve, induction therapy may be given first, followed by reassessment. Understanding the relationship between the tumor and surrounding organs earlier allows doctors and patients to prepare more thoroughly.

AI brings another way of exploration: simultaneously calculating the texture and shape of the lesion and the three-dimensional relationships with surrounding tissue, learning complex combined features from preoperative CT and postoperative outcomes. Information that doctors find difficult to compare item by item can thus be analyzed and tested in batches. This study, based on iMedImage and combined with radiomics, carried out staging prediction in 323 cases of thymic epithelial tumors, with 258 cases for training and 65 cases for internal testing. The model distinguished the two groups of Stage I-II versus Stage III-IV, with a test accuracy of 95.38% and an AUC of 0.9328. The results were published in the European Journal of Radiology Open in July 2026. (From postoperative confirmation to preoperative prediction, staging task accuracy corresponds to 65 internal test cases. Source: DIAGENS 2026 research paper and clinical literature.) The significance of this work is to extract new information from existing clinical CT scans that can help preoperative judgment. Once clinically recognized in the future, it is expected to add evidence for assessing resectability, preparing for surgery, and discussing treatment sequencing.

The same study also used all 659 cases to distinguish thymic epithelial tumors from other lesions, and conducted WHO pathological risk grouping on 323 of those thymic epithelial tumor cases. A single set of CT data supports continuous inquiry into lesion nature, degree of invasion, and histological risk. Research of this kind is not easy. A 2023 study retrospectively reviewed ten years of cases from three hospitals, with contrast-enhanced CT analysis including 373 people; another 2024 study of 187 cases required about 30 minutes per case for segmentation and slice-by-slice correction. Behind several hundred cases lies long-term accumulation and a large amount of professional labor. With the same CT scan, AI brings more information into computation and more medical questions into research.

Breast cancer recurrence prediction pushes the research toward the future. What patients care about most after surgery is whether the cancer will come back. Patients with the same stage may have different long-term outcomes. Pathological slides show cell and tissue morphology, ultrasound shows another side of the tumor, and clinical data record the patient's condition. AI puts this information into the same framework and systematically compares the combinations of a large number of subtle features with subsequent outcomes, giving researchers a chance to find individual differences beyond traditional grouping. According to DIAGENS' interim results roadshow information, the company has conducted a 673-case study combining pathology, ultrasound, and clinical information, using AI to find links between subtle features and recurrence risk. The optimized model's five-fold cross-validation C-index was 0.76, which evaluates risk ranking ability. Once clinically recognized in the future, such tools are expected to help doctors assess individual risk in greater detail and support long-term management and risk communication.

From understanding current lesions to predicting future outcomes, AI4S is turning more medical questions into computable, testable research. The key for DIAGENS lies in the common foundation supporting these studies. iMedImage is first pretrained on multiple types of imaging, bringing transferable features to new topics; iMedLoop organizes data processing, annotation, training, evaluation, and delivery, reducing the repeated investment of each team developing from scratch. This can be understood as "building a common foundation first, then doing specialized training." The several hundred cases in a new topic mainly carry the learning of a specific task; the imaging features accumulated previously have already become its starting point. Doctors can then focus more energy on raising questions, judging results, and designing the next round of research.

According to data disclosed in the company's annual report, vertical model development can use as few as about 200 images, with a development cycle of 2 to 3 months. About 15 months after iMedMaaS was launched, as of the end of June 2026, a total of 158 model projects had been carried out, cooperating with 99 hospitals and covering 61 disease directions. (A common foundation lowers the development threshold for new topics.) When the R&D threshold falls, limited budgets have a chance to support more topics. What changes is not only research speed, but also whether a medical idea can become reality. In the first half of 2026, DIAGENS' model service revenue was about 94.54 million yuan, up 101.1% year on year, also providing a commercial fulcrum for continuous R&D.

For investors, this means a process that can continue to expand: existing capabilities serve current customers, the common technical foundation supports more topics, and new research provides sources for future deliverable models and services. This process is still moving forward. In September 2026, DIAGENS and The Hong Kong Polytechnic University established a joint laboratory for general artificial intelligence and medical applications, focusing on medical imaging, medical foundation models, and automation technology, connecting basic research with clinical translation. Pushing this path further forward, DIAGENS' imaginative space may lie in the embryonic form of a "medical world model." A world model learns how the environment changes and how actions affect outcomes; the vision of a medical world model is to infer disease progression based on patient status and further explore the impact of different interventions. The thymus study learns the disease states corresponding to imaging, while the breast study explores the future risk indicated by current features. If these capabilities enter application and continuously connect with follow-up, treatment, and real outcomes, there is a chance to move toward disease course simulation; when models can reliably learn the relationship between interventions and outcomes, doctors may be able to preview the changes a same patient might experience under different treatment sequences. The most moving imagination of a medical world model is to let one real diagnosis and treatment decision receive more rehearsals in computation. Diagnosis provides the current state, follow-up tests previous predictions, and real outcomes drive model improvement. If such a cycle forms in the future, more questions that are difficult to test repeatedly in reality will have a chance to be explored in computation first.

As Kevin Kelly, the "godfather of Silicon Valley," proposed in a conversation with Song Ning, chairman and CEO of DIAGENS, a lasting asset may be the process of continuously generating new medical AI. DIAGENS' models and platform are providing the foundation for such a process. Its scarcity in AI4S may lie precisely here: turning individual medical questions into new capabilities, and then advancing toward a larger system for understanding and inferring disease.