Zhitong Finance App News, Deshi-B (02526) announced that the Group continues to cultivate artificial intelligence in medical imaging, independently develop a large iMediimage® base model, complete the full-modal architecture upgrade in the first half of 2026, and launch the iMedloop™ medical imaging AI research and production acceleration platform in July of the same year. Relying on large-scale, high-quality medical imaging data accumulated by the Group, the medical imaging understanding and analysis capabilities of the Group's base model have been continuously improved through technological innovation, continuous training and expansion of the training scale through underlying algorithm technology. The board of directors of the company is pleased to announce that the Group's latest generation medical imaging models, iMED-6.0 Thoth (Thoth) and iMED-6.0 Aegis (Aegis), have made significant technical progress. The two are medium parameter scale and small parameter scale versions of the iMED-6.0 series, respectively.
The Group's iMED-6.0 series models achieved world-leading results in the world-renowned large-scale multidisciplinary multi-modal understanding and reasoning evaluation benchmark (“MMMU (Test)”) and the full-modal large model evaluation platform MedBench, a medical AI professional evaluation platform. MMMU (Test) covers six major fields and 30 categories, including medicine, science and engineering. According to the group's test results, Thoth's overall score in all fields is higher than that of other models. Thoth and Aegis ranked first and second in the world in the “Health and Medicine” category, showing their excellent multi-modal understanding and reasoning ability in medicine, and further reflecting their potential for generalization across disciplines and multiple tasks. According to the evaluation results obtained after the group model was submitted to MedBench, the overall score of Thoth and Aegis was also higher than the highest score of the model already listed in the Full Modal Big Model List during the same period, ranking first and second in the world. Among them, Thoth scored 88.8 points in the “Medical Visual Perception and Text Extraction” category, exceeding the previous high score of 20.3%.
Based on the general computed tomography (CT) vertical model developed by Aegis, the Group has made important technical progress in comprehensive analysis of CT images with multiple organs and multiple abnormalities. The model supports chest and abdominal scan and enhanced CT, and can comprehensively analyze multiple organs and multiple types of abnormalities within a single scan, covering 509 diseases and imaging signs. In the group's evaluation of nearly 40,000 CT images and reports from multiple hospitals, the average area under the curve (AUC) (index of ability to identify abnormalities, the closer to 1) of the 249 key tasks selected by the group reached 0.930. As a reference to the public study, the RADAR study published in “Science” revealed 146 abdominal imaging findings, and the average AUC verified at various centers was 0.895. Aegis averaged an AUC of 0.855 across all 509 tasks.
In order to further enhance its core technical capabilities, the Group believes it is essential to establish more objective, comprehensive, and difficult professional evaluation benchmarks close to clinical practice. The Group is continuously upgrading its self-built DoctorBench professional evaluation benchmark to further accurately and comprehensively measure the model's ability to understand, reason and comprehensively apply medical images in complex scenarios for challenging medical tasks. The Group plans to announce the results of a new round of evaluation in the future, including the relevant performance and R&D progress of the iMED-6.0 Horus model with the largest parameters of the Group.
The board of directors believes that the continuous improvement of base model capabilities is an important driving force for enhancing the core competitiveness of the Group's “model factory” and expanding commercialization space. A stronger base model is expected to improve iMedloop™'s model development and training efficiency, speed up the supply of specialist models for different organs, diseases and application scenarios, and support medical experts to speed up research and verification of new methods. At the same time, relevant technological advances are expected to accelerate the deployment and application of diagnostic models to primary care systems, help medical institutions improve the efficiency and accuracy of image reporting through inference services, and drive the expansion of case-by-case model reasoning services. As supply models and service scenarios increase, the Group is expected to expand customer coverage and case service scale, and promote the transformation of technical advantages into continuous commercial value. The Group will continue to explore the maximum capacity of the pedestal model, strengthen training and professional evaluation, and strive to accelerate the intelligence of global medical imaging and improve human health.