In our paper Tissue-Type Mapping of Gliomas by Raschke et al we are developing a method of automatically labelling the various tissue types found in gliomas from magnetic resonance images (MRI). Gliomas are the most common primary brain tumours in adults and grow by infiltration into normal brain tissue. Detection of the infiltration boundaries and whether the tumour is a slow-growing low-grade, or a malignant high-grade, is important for planning optimal treatment. Better labelling and delineation of the various tissue types in this heterogeneous tumour could aid in surgical and radiotherapy planning.
MRI provides a variety of imaging modalities that neuroradiologists can use to aid their diagnosis of the tumour grade and delineation of its margins. In our study we are using Bayesian statistics to combine several MRI image types and create a tissue-type map as shown below. Standard clinical scans include the T1-weighted image acquired after injecting a gadolinium-based contrast agent, in which high signal indicates areas where the blood-brain barrier is leaky as found in high-grade gliomas. In the FLAIR image, high signal indicates areas of increased water content, which may be oedematous brain or areas of tumour infiltration. The p image is a quantitative measure of water diffusion and relates to how structured the tissue is. High p occurs in necrosis and oedema. In addition we acquire q diffusion images, which is high in areas of intact white matter, and decreases in areas of tumour infiltration and in oedema, as well as PD- and T2-weighted images which show an increase as tissue water increases and the brain becomes less structured due to the tumour.
All these images vary in subtly different ways according to the tissue structure and we use another technique, MR spectroscopy (MRS) to determine the metabolites present in the tissue, so that we can label small regions as low-grade, high-grade, oedema, necrosis or normal brain tissue. Finally we derive a Bayesian statistical algorithm that recombines all the different MR images to create tissue probability (b) maps and color-coded tissue type maps (d). Once the algorithm has been derived we no longer need to use the MRS and can create whole-brain tissue-type maps as shown for a single imaging slice below (e).
The tissue-type map in (e) could help determine where best to perform a biopsy to assess the tumour tissue in more detail with genetic profiling, or where best to apply the highest dose of radiotherapy and which areas to minimise the dose to avoid affecting the more normal tissue.
This methodology is currently being further developed and validated with an Innovate UK grant.
