Secure, browser-based AI that speeds up clinical brain-scan analysis

What’s the problem? 

Analysing a brain scan often means outlining lesions by hand — slow, painstaking work that can vary from one clinician to the next. AI can speed this up, but the tools usually run in the cloud, so sensitive patient images have to leave the hospital, while adapting a model to a particular clinic’s patients normally demands specialised GPU hardware and technical expertise that most clinical teams don’t have. 

What’s UQ’s innovation? 

Dr Thuy Dao and her team from the School of Electrical Engineering and Computer Science have developed Deepsyence, which brings AI-assisted bran-scan analysis into an ordinary web browser. It uses interactive foundation models like MedSAM and EfficientSAM to rapidly segment lesions. Clinicians can seamlessly refine these segmentations to fine-tune models locally without compromising patient data. 

What’s the impact? 

Because Deepsyence runs in the browser on ordinary hospital computers — no GPU required, and no scan ever leaving the machine — the tool fits within the tight software and privacy constraints of real clinical settings. The tool was demonstrated on scans from patients with aphasia, a language disorder caused by damage to the brain arising from stroke, injury, infections and other conditions. Finetuning the classification model on a small set of local cases lifted its accuracy at classifying aphasia types from 54 to 73 per cent, showing how clinicians could hone the tool on their own patient populations rather than relying on a fixed, one-size-fits-all model.

Led by 

• Dr Thuy Dao 
  Platform Engineer, School of Electrical Engineering and Computer Science, UQ 

UQ Project Team

• Dr Steffen Bollmann
  School of Electrical Engineering and Computer Science 
   
• Xincheng Ye
  School of Electrical Engineering and Computer Science
   
• Dr Michèle Masson-Trottier
  School of Electrical Engineering and Computer Science 

99 more AI innovations at UQ