Breakthrough in AI-driven Diagnostic Imaging by CMU Health Cluster

Researchers at Chiang Mai University have unveiled a diagnostic imaging model that brings clinically deployable accuracy to early-stage anomaly detection. Built jointly by the Digital & AI and Health clusters, the system has been validated across four hospitals in Northern Thailand and consistently outperforms general-purpose vision baselines.

The architecture pairs a domain-tuned vision transformer with a calibration layer that adapts to scanner-level variation, a long-standing barrier to deploying machine learning across hospital networks. In the validation cohort, the model achieved 94.3% sensitivity and 88.1% specificity at the operating point preferred by reviewing radiologists.

Beyond accuracy, the team focused on workflow fit. The inference pipeline runs locally on existing radiology workstations, never sends pixels off-site, and surfaces region-level evidence that radiologists can confirm or override. An open-source toolkit ships with regional weights so that affiliated hospitals can re-tune to their own patient population without sharing raw imagery.

The next phase of the project, supported by a multi-year grant from the National Research Council, will extend the system to ultrasound and pediatric MRI cohorts and add a longitudinal layer that tracks subtle changes between visits.