Automatic Brain Tumour Segmentation

Doctors must locate Braint tumours

• Segment Magnetic Resonance (MR) images (seperate tumour from non-tumour)

Typically done by hand

• labout intensive and often inaccurate/inconsistent

We developed automated segmenter

• program that automatically separates tumour from normal, without human input
• program was learned from prior segmented images

More accurate than other automated tumour segmentation algorithms

Brain Tumour Growth Modelling

Brain tumours have

• regions VISIBLE in MRI scans and
• regions INVISIBLE (radiographically occult)
• Treating ONLY visible region is not sufficient!
• Currently: doctors treat VISIBLE + 2cm border
• But... kills some healthy tissue, spares some cancer cells!
   

Implemented system that predicts how brain tumours will grow
based on scan + properties of the patient, tumour

• Learned from MRI scans of earlier patients
• More accurate than other growth predictors
   

How do our Systems Work?

• Segmentation based on two approaches:
• 1. Probabilistic approach - conditional random fields
• 2. Variational approach - level sets
• Both approaches involve training on 100's of expert-labeled patient scans.
• Growth prediction: Incremental "growth" process, trained with time sequence
• of labelled MR images for many different patients

Current Projects

• Better accuracy using improved Machine Learning algorithms
• Fast (if approximate) algorithms
• New modalities: Diffusion tensor images, Magnetic resonance spectroscopy
• Database of previous labelled images, for fast recall and use
• Translate from prototype to working, deployed system