DIAGENS makes a CT AI model for 509 diseases and findings
In its test on almost 40,000 CT exams, the model had a score of 0.855 across all 509 targets.
DIAGENS, a company on the Hong Kong stock exchange, announced results for iMed-6.0, its medical imaging AI model. It also made a CT model for 509 diseases and imaging findings in chest and abdomen scans. The company says the score is 0.855 on AUC, a score for how correctly a model sorts exams. The test used almost 40,000 exams from some hospitals.
How it works
The CT model reads scans of the chest and abdomen, with and without contrast dye. It sorts exams into the 509 targets with no more training. iMed-6.0 also reads 3D images, tissue slides and images from different times. For some tasks, the output shows the slices or regions in the source images.
The numbers
- AUC on 249 targets from public benchmarks: 0.930.
- MedBench, a test by Shanghai AI Laboratory: Thoth and Aegis were numbers 1 and 2, with 64.4 and 61.7.
- MMMU Test: Thoth had 66.2%, above the 65.9% of the best result before.
What is not known
The 0.855 and 0.930 scores come from a test by the company. The source names no external check of the CT model. It gives no data on use in a clinic.
Insight: A tool seller gains when each hospital makes custom models
DIAGENS sells zero-code tools, and hospitals can train models for their tasks. It says it did 158 model projects with 99 hospitals. A test in October 2026 found that a small tuned network had a higher score than 2 large language models on skin lesions. This reads as a move to hold the procedure for new models, and the vendor with the hospital data can possibly set the standard.
Sources
Posted