DIAGENS-B(02526) The scarcity of AI for Science may be hidden in the prototype of a medical world model.

Title context: DIAGENS-B(02526) The scarcity of AI for Science may be hidden in the prototype of a medical world model. Text: Before a tumor surgery begins, what a doctor most needs is to know as much as possible about what they will face. 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 bear on the surgical plan, yet reliable answers often have to wait until postoperative pathological examination. The thymus study in which Diagens (02526) participates 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 manually one by one, testing the connections between imaging and disease, and making more medical questions affordable and faster to answer. Specifically in this study, the MasaokaKoga 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 permits 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 exploring: simultaneously computing the texture and shape of the lesion and its 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 is based on iMedImage, combined with radiomics, and conducted staging prediction in 323 cases of thymic epithelial tumors, with 258 for training and 65 for internal testing. The model distinguished between Stages III and Stages IIIIV, 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 tests. Source: Diagens 2026 research paper and clinical literature) The significance of this work is to extract new information from clinically available CT that can help with preoperative judgment. Once clinically validated 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 the thymic epithelial tumor cases. One set of CT data supported 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 great deal 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. After surgery, what patients care about most is whether the cancer will come back. Patients with the same stage may have different long-term outcomes. Pathology slides show cell and tissue morphology, ultrasound shows another side of the tumor, and clinical data record the patient's situation. AI puts this information into the same framework, systematically comparing the combinations of large numbers of subtle features with subsequent outcomes, giving researchers the opportunity to find individual differences beyond traditional grouping. According to Diagens' interim results roadshow information, the company has conducted 673 studies combining pathology, ultrasound, and clinical information, using AI to find connections 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 validated in the future, such tools are expected to help doctors assess individual risk more carefully 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 to 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 the specific task; previously accumulated imaging features have already become its starting point. Doctors can thus focus more energy on raising questions, judging results, and designing the next round of research. From the 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. iMedMaaS has been online for about 15 months. As of the end of June 2026, it had cumulatively carried out 158 model projects, cooperated with 99 hospitals, and covered 61 disease directions. (A common foundation lowers the development threshold for new topics) When the R&D threshold declines, limited budgets have the opportunity to support more topics. What changes is not only the speed of research, but also whether a medical idea can become reality. In the first half of 2026, Diagens' model service revenue was approximately RMB 94.54 million, a year-on-year increase of 101.1%, also providing a commercial fulcrum for continued R&D. For investors, this means a process that can continue to expand: existing capabilities serve current customers, the common technology foundation supports more topics, and new research in turn provides a source 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 technologies, connecting basic research with clinical translation. Pushing this path further forward, Diagens' imaginative space may lie in the prototype 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 simulate disease progression based on patient status and further explore the effects 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 an opportunity to move toward disease progression simulation; when models can reliably learn the relationship between interventions and outcomes, doctors may be able to rehearse the changes a single patient might experience under different treatment sequences. The most compelling imagination of a medical world model is to let one real clinical decision be rehearsed many times 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 the opportunity to be explored first in computation. As Kevin Kelly, the "godfather of Silicon Valley," proposed in a conversation with Song Ning, Chairman and CEO of Diagens, a durable asset may be the process of continuously generating new medical AI. Diagens' models and platform are providing the foundation for such a process. Its AI4S scarcity may lie precisely here: turning one medical question after another into new capabilities, and then advancing toward a larger system for understanding and simulating disease.