Skip to content
AI.info

The Pulse

DIAGENS introduces CT model designed to identify 509 imaging targets

DIAGENS introduced a model designed to identify diseases and imaging findings in chest and abdominal CT scans. The company also reported internal validation results and benchmark scores for its iMed-6.0 models.

DIAGENS introduces CT model designed to identify 509 imaging targets

A new DIAGENS model is designed to identify 509 diseases and imaging findings in chest and abdominal CT scans. The company introduced it on November 12 alongside benchmark results for iMed-6.0, the latest generation of its iMedImage medical imaging foundation model.

Source accessnewswire.com

The model is designed for both non-contrast and contrast-enhanced examinations of the chest and abdomen. It covers 263 chest targets and 391 abdominal targets, with overlapping targets deducted from the total.

Internal validation

DIAGENS said an internal validation used nearly 40,000 CT examination-and-report pairs from multiple hospitals. Across all 509 targets, the model achieved a macro-averaged AUC of 0.855 for zero-shot classification.

For a subset of 249 key targets represented in internationally available public benchmarks, the macro-averaged AUC was 0.930. The CT model was developed using iMed-6.0 Aegis, one of the models in the company's foundation-model series.

Benchmark results

DIAGENS also reported scores for its Thoth and Aegis models in MedBench's multimodal API evaluation, where it said they ranked first and second overall and in each of three core categories. MedBench was developed by Shanghai AI Laboratory and is part of the OpenCompass evaluation ecosystem.

The two models scored 64.4 and 61.7 overall, respectively. The categories covered medical visual perception and text extraction; cross-modal semantic understanding and reasoning; and clinical decision support and reasoning.

In medical visual perception and text extraction, Thoth scored 88.8 and Aegis scored 84.9. Those scores were 20.3% and 15.0% higher, respectively, than the highest score of 73.8 among the other evaluated models, according to the company.

The company said Thoth also achieved 66.2% overall accuracy on the Massive Multi-discipline Multimodal Understanding and Reasoning benchmark, above the previously published best result of 65.9%. The benchmark spans six broad domains and 30 subjects. In its health and medicine category, Thoth and Aegis scored 74.9% and 71.1%, respectively, placing first and second in the evaluation.

Development tools and business

The underlying iMed-6.0 model processes 3D imaging, whole-slide pathology and temporal imaging data. It supports image interpretation, anatomical and lesion analysis, and report drafting and review; for supported tasks, outputs can be linked to slices or regions of the source images.

DIAGENS combines those capabilities with iMedLoop, a development platform covering data preparation, training, evaluation, deployment and feedback. Its iMedMaaS offering lets clinicians and hospitals train, deploy and scale custom models through a zero-code interface.

As of June 30, 2026, the company said it had undertaken 158 model projects and established research and development partnerships with 99 hospitals. Its products and services had reached more than 400 healthcare institutions.

For the six months ended June 30, DIAGENS reported revenue of RMB 108.7 million, up 21.0% year over year. Model-as-a-Service revenue reached RMB 94.5 million, up 101.1% year over year and accounting for approximately 86.9% of total revenue.

Sources

Related stories

Explore

More articles