• Automatic Brain Tumour Segmentation
• Tumour Growth Prediction

Here we consider segmenting tumours based on axial MR slices, using the T1, T1c, and T2 modalities.
See Segmentation page for an overview of our approach.
See here for information about possible licensing opportunities.

We consider two training regimes:

• The easier case is "intrapatient training": Here, a human expert provides labels for a few of the slices of the patient's brain. These manually labelled images are then used as training input to our learning algorithm. The resulting classifier is then used to produce labels for other unlabeled (axial MRI) slices within the brain of the same patient. We believe that our system has solved this "intrapatient segmentation" task.
• By contrast, "interpatient training" means training on the images from one set of patients, then testing on images from a novel patient.

Our team has implemented a pipeline (based on (Schmidt, 2005)) that can run automatically to produce segmentation labels for both intra-patient studies and inter-patient studies. In either case, our system receives the set of manually labelled images as training input to our pipeline, which uses that information to learn the relevant parameters. (We developed the "OSP" software for helping the expert provide these labels.)

The resulting trained classifier can then be used to produce labels in unseen slices of the brain (of the same brain in the intrapatient case, or of new brains for interpatient case).

Using this data we have tested many different ways of segmenting tumours. Each method has its own advantages. Some of our methods are optimized for speed, while others are optimized for accuracy. Our results indicate that the most accurate classification algorithm is support vector machines. The results also indicate that our optimal feature set should contain features including, but not limited to, symmetry, intensity values, template information, multiresolution images, and brain priors.

Figure 1 shows our results come very close to matching the expertly segmented labels exactly. Moreover, another advantage of our method is that it is dynamically learned, which means that it can be used to learn to segment alternative regions of interest. For example if the physician is interested in edema the tool can be used to segment regions of edema. Alternatively if the physician is interested in enhancing tumour area, the tool can be used to segment only this region. (We are now beginning to use it to segment anatomical features, such as the brain stem and eyes.)

Below are the Jaccard scores based on our latest tests:

• Intrapatient
  • Enhancing Tumour: 0.859
  • Tumour + Edema: 0.796
  • Gross Tumour Volume: ?
• Interpatient
  • Enhancing Tumour: 0.73
  • Tumour + Edema: ?
  • Gross Tumour Volume: ?

Our experiments confirm that it is easiest to segment the enhancing tumour area; this is expected as this area is the most visible on the MRIs. The 'tumour + edema' region is more challenging, as it is not always as easy to see in the MR images. The most challenging is gross tumour volume, which is often very difficult to see visually -- indeed, many times the only way for a physician to determine gross tumour volume is by means of biopsy.

Results

• One of the world's best tumour segmentors, based on T1, T2, and T1c
• First to fully automatically recognize and segment non-trivial cases, including
  • heterogeneous content
  • NO enhancing area
• Accurate enough to automate radiation therapy target planning (ie, provide initial volume, which oncologists can then modify -- eg, by adding 2 cm to the boundary)
  ... a huge time-saving for physicians
• First (and only!) system that addresses ...
  • inter-slice variations
  • gain field
  • inter-patient/machine problems
• Meaningfully combine 'unique' contributions of most of recent influential results ... via Machine Learning
• Provisional patent!

We conclude from our results that, though glioma growth prediction is a rather challenging problem, it is feasible to model glioma growth based on learning and classification more accurately than with standard methods. Several scenarios in the experiments showed that our classification-based diffusion model (CDM) can `track' glioma diffusion patterns, in particular when the enhancing tumour spreads along the edema regions.

Tumour growth prediction is a more difficult task then tumour segmentation according to our experiments. We also conclude that glioma growth does not conform with radial uniform growth (which is implicitly assumed by standard treatment strategies), but tumours tend to grow into asymmetric volumes. In addition, several factors are involved in this diffusion process, and it is not sufficient to account only for the heterogenous brain tissue when modeling tumour growth. To successfully model glioma growth, it is important to incorporate information specific to the voxels adjacent to the tumour, the edema regions, the brain anatomy and to the patient.

CDM performs more accurately in most scenarios as compared to radial uniform growth (UG). It is also more accurate than diffusion based on grey and white matter (GW) in several scenarios. Currently, CDM's prediction is mainly based on features extracted from the MRI scans and can only be as accurate as the threshold of detection of tumour regions with MR imaging and the segmentation of tumour volumes.