Tissue-type mapping of gliomas

Felix Raschke, Thomas R. Barrick, Timothy L. Jones, Guang Yang, Xujiong Ye, Franklyn A. HoweNeuroImage 2018 · DOI

TL;DR

This paper introduces a novel multimodal MRI (mMRI) technique for creating detailed "tissue-type maps" of gliomas. By combining information from various MRI sequences and using a Bayesian approach, the method can differentiate between healthy tissue, low-grade tumors, high-grade tumors, necrosis, and edema, offering a more precise spatial understanding of tumor heterogeneity than traditional methods. This has the potential to significantly improve treatment planning and patient management for glioma patients.

What Problem Does This Paper Address?

Gliomas, the most common primary brain tumors, are notoriously heterogeneous. This means they aren't uniform masses but rather complex mixtures of different tissue types, including viable tumor cells (varying in grade), necrotic (dead) tissue, and surrounding edema (swelling). Accurately mapping these different components is crucial for effective treatment. Current methods, like standard MRI, can indicate tumor grade but struggle to precisely delineate the infiltration margins – how far the tumor has spread into healthy brain tissue. This imprecision leads to generalized treatment margins in radiotherapy, potentially overtreating healthy tissue or undertreating the actual tumor extent. Histopathology can provide definitive grading but is invasive and only samples a small area. This paper aims to bridge this gap by developing a non-invasive, whole-brain method to create detailed maps of tumor tissue types.

Methods

The core of this work is a "tissue-type mapping" algorithm that leverages multimodal MRI (mMRI) data. Here's a breakdown of their approach:

  • Multimodal MRI Acquisition: They acquired several MRI sequences for each patient: T1-weighted (with contrast), FLAIR, T2-weighted (T2w), Proton Density-weighted (PDw), and Diffusion Tensor Imaging (DTI). DTI data was processed to derive maps of isotropic (p) and anisotropic (q) diffusion.
  • ¹H MRSI for Ground Truth Labeling: Crucially, they used ¹H Magnetic Resonance Spectroscopic Imaging (MRSI) to obtain metabolic information from specific voxels. This MRSI data was used to objectively label regions as "pure" low-grade (GII) or high-grade (GIV) tumor. This objective labeling is a key advantage over subjective manual labeling.
  • Bayesian Classification: They developed a Bayesian framework to combine the information from the different MRI modalities. This framework uses "a priori probability density distributions" (PDDs) for each tissue type (healthy brain, GII, GIV, necrosis, edema). These PDDs were derived from the MRSI-labeled regions and the corresponding MRI signal intensities. Bayes' theorem was then applied to each voxel to calculate the posterior probability of it belonging to a specific tissue class.
  • Superpixel Segmentation: To reduce noise and smooth the resulting probability maps, they employed superpixel segmentation. This technique groups image voxels with similar characteristics into larger, more manageable regions, allowing for local averaging of tissue probabilities.
  • Tissue-Type Color Maps: Finally, the probability maps were translated into visually intuitive color maps. Different colors represent different tissue types (e.g., red for high-grade, green for low-grade/infiltrated, blue for necrosis, black for edema), providing a clear spatial representation of tumor composition.

What's novel here is the use of MRSI for objective tissue labeling to build the prior PDDs for a Bayesian classifier, combined with superpixel segmentation for robust mapping, all applied to a multimodal MRI dataset for whole-brain coverage.

Datasets

The study utilized data from several sources:

  • Glioma Patients: 25 adult patients with histologically confirmed gliomas (WHO grades II, III, and IV).
  • Brain Metastasis Patients: 10 patients with brain metastases, used to characterize vasogenic edema.
  • Healthy Controls: 10 age-matched individuals, used to characterize normal brain tissue (grey matter, white matter, CSF).

The MRSI data from the 25 glioma patients had been previously analyzed, and this study leveraged those results for labeling. The MRI data for healthy controls was sourced from the St. George's Neuroimaging in the Elderly (GENIE) study.

Key Insights and Ideas

  • MRSI as an Objective Labeler: Using ¹H MRSI to objectively identify "pure" tumor tissue for training the classification model is a significant improvement over subjective manual labeling, which can be prone to bias and limited in scope.
  • Multimodal Data Synergy: Combining multiple MRI sequences (diffusion, PD, T2) in a Bayesian framework allows for better discrimination between tissue types than any single modality alone. The paper highlights how different modalities contribute unique information (e.g., diffusion for infiltration, PD/T2 for tissue characteristics).
  • Bayesian Approach for Uncertainty: The Bayesian framework naturally provides probability maps, reflecting the uncertainty in classification. This is more informative than a hard classification and allows for flexible thresholding to define different tissue boundaries.
  • Superpixels for Robustness: Superpixel segmentation helps to smooth out noise and create more anatomically plausible tissue boundaries, leading to more reliable maps.
  • Clinical Relevance of High-Grade Volume: The study found a strong correlation between the automatically segmented high-grade tumor volume and patient survival, suggesting this mapping technique could have direct prognostic value.

Key Results

  • Lesion Volume Correlation: Automatically segmented total lesion volumes correlated well with manually delineated volumes (Pearson's r = 0.87, p < 0.001).
  • Survival Prediction: Patients with shorter survival times (< 2 years) had significantly larger high-grade tissue volumes compared to those with longer survival times.
  • Tissue Type Differentiation: The method successfully differentiated between normal brain, vasogenic edema, low-grade tumor, high-grade tumor, and necrosis.
  • MRSI vs. mMRI Comparison: Tissue proportions estimated by the mMRI tissue-type technique showed good correlation with ¹H MRSI analysis (Distance Correlation r = 0.79-0.84).
  • Boundary Agreement: Good correspondence between automatically and manually segmented lesion boundaries was achieved (Dice coefficient 0.77 ± 0.13).

Limitations and Future Work

  • Retrospective Study & Small Dataset: The study is retrospective and based on a relatively small cohort (25 glioma patients), necessitating validation with larger, prospective datasets.
  • WHO Classification Update: The study used the older WHO classification system. Future work should incorporate the 2016 WHO classification, which includes genetic subtypes that may have distinct MRI characteristics.
  • Semi-Quantitative MRI: The use of semi-quantitative PDn and T2n images could be improved by using fully quantitative methods, such as MR fingerprinting, for more precise tissue characterization.
  • Optimal Thresholding: The optimal probability threshold for defining tumor infiltration requires further investigation through longitudinal studies.
  • Integration of Other Modalities: The Bayesian framework is expandable and could readily incorporate other quantitative imaging methods like DSC-MRI to further enhance classification accuracy.
  • Clinical Application: While promising, the clinical applicability needs further validation, particularly in guiding targeted biopsies, surgical planning, and radiotherapy dose painting.

Summary updated

Abstract

<sup>1</sup>H MRSI can label tumour tissue types to enable development of a mMRI tissue type mapping algorithm, with potential to aid management of patients with glial tumours.
BibTeX
@inproceedings{raschke2018tissue,
  author = {Felix Raschke and Thomas R. Barrick and Timothy L. Jones and Guang Yang and Xujiong Ye and Franklyn A. Howe},
  title = {{Tissue-type mapping of gliomas}},
  booktitle = {NeuroImage},
  year = {2018},
  doi = {10.1016/j.nicl.2018.101648},
}
Figure