In this project, we use machine learning to study and model glioma growth. Given a tumour observed on an MRI scan of a patient's brain, we try to predict where the tumour will grow, which will help oncologists direct radiation treatment.

For this purpose, we have developed a 3D classification-based diffusion model (CDM) that learns from existing patients and predicts tumour growth at the voxel level for a patient based on features specific to the patient and the tumour, and attributes of the voxel.

Motivations for Glioma Growth Modeling:

Glioma growth modeling has direct application in radiotherapy for the treatment of primary brain tumours. The conventional method of treating gliomas with conformal radiotherapy is to irradiate both the visible tumour mass observed on the MRI scan as well as a 2cm margin which typically includes both healthy tissue and malignant tendrils. This radial 2cm margin does not take into account patient or tumour information but simply assumes that glioma cells have uniformly infiltrated through surrounding healthy tissue. This treatment method often compromises brain function and could still leave behind cancer tendrils allowing tumour recurrence after a number of months. Therefore, it has become necessary to find ways to model glioma diffusion more accurately, which will help radiation oncologists target the diseased regions more properly.

Problem Definition:

In this project, we address the problem of glioma growth modeling as the task of predicting the size and shape of the tumour at a later time (e.g., six months later) based on attributes of the visible tumour mass currently detected on the MRI scan, and attributes of the regions near the tumour. Our approach assumes that cancer cells and tendrils that infiltrated in the nearby healthy tissue but remained unobserved on the MRI scans are responsible for further glioma diffusion and recurrence.

Project Framework:

The framework of the project consists of the following main steps, also provided through a walk-thru demo of the OSP/ASV program:

• Noise reduction: reducing effect of noise, intensity inhomogeneity in the MRI scans.


• Registration: spatially aligning the image to a template in a standard coordinate system and warping the image regions to the template.


• Intentensity standardization: scaling the image intensities to provide more meaningful approximations of the image regions aligned to template.


• Tumour segmentation: recognizing the tumour mass and its boundaries on the MRI scan either manually (drawing the contour of the tumour on the image) or automatically.


• Tissue segmentation: separating regions of the brain scan into two main tissue types, white matter and grey matter, as well as cerebrospinal fluid (CSF).


• Feature extraction: identifying voxel features that provide information about regions near the tumour.


• Classification: using a classifier, learned from existing patient data and features of the voxels, we assign each unlabeled voxel in the brain volume of a new patient to either tumour or non-tumour class.


• Diffusion modeling: simulating the tumour growth, given an initial tumour volume of a patient at one time, based on the labels predicted through classification, produces the predicted shape of the tumour in 3D at a later time when the tumour has reached a particular volume.

Our Machine Learning Approach:

Our approach consists of a 3D modeling system at the voxel level based on learning a classifier from existing patient data (i.e., MRI scans of a number of patients) and using the learned classifier to predict whether the voxels of a new patient are healthy or cancerous.

Our performance system is initialized with a tumour (the gross tumour volume, GTV) of the given patient, at one time point, and the system models the shape of the tumour when it is v voxels larger (see Figures 1 and 2). The system uses a diffusion model to predict the tumour growth (i.e., where the additional v voxels will be).

In order to predict these additional voxels, we first learn a classifier (e.g., Naïve Bayes, Logistic, or SVM) from a time series of MRI scans based on a set of pairs of labeled brain volumes of various patients; each pair is obtained from the same patient at different time points. The learning task includes a number of attributes about each voxel in each brain volume, as well as a class label: tumour or non-tumour. Training data consists of the voxels adjacent to the tumour and that a perfect diffusion algorithm would consider in growing from the patient's initial tumour at one time to the final tumour volume at a later time.

Current Results:

Current results are based on manually labeled pairs of tumours for 17 patients. In training, we use attributes from the pairs of tumour volumes for 16 of the patients, and we use the learned model to predict tumour growth for the 17th patient (see Figure 1).

A diffusion algorithm uses the voxel labels (tumour or non-tumour) predicted by the classifier along with a probability threshold to eventually assign each voxel to either class, as shown in Figure 2.

Future Work:

Future directions include improvements and additions at each of the data, the framework, and the machine learning levels:

• Previous experiments were performed on a small patient set (where tumours have been manually labeled), and it is therefore essential to expand the training dataset using automatically segmented tumour volumes.


• Other types of data could be used, e.g., PET, and MRS, as opposed to solely using MRI scans.


• Improvements at the framework level consists of using better image processing techniques to prepare the MRI scans for the feature extraction, learning, and classification tasks.


• Features that could be added include metabolic and genetic data as well as anatomical attributes of the brain and other tumour-specific features. Also, automatic feature selection can be used on both existing and new features to find the set of features that contains the most important information.


• Improvements on the machine learning level include experimenting with a variety of classifiers and exploring their performance with different feature sets.