You are here: Machine Learning for Precision Health»Multi-Parametric MRI and Big Data Analytics

Multi-Parametric MRI and Big Data Analytics

Multi-Parametric MRI for Improved Cancer Diagnostics and Prognostics

Most pathology semi-automated quantification tools provide only rudimentary, often inaccurate pathology measurements and are limited to analyzing a single modality.However, to detect and label amorphous pathology such as (1) the tumors and edema in cancer, (2) contusions, hemorrhages, and microbleeds in traumatic brain injury, and (3) white matter lesions in diseases such as Multiple Sclerosis, the fusion of multiple sources of information is required.To overcome the limitations of existing tools, we developed a flexible machine learning-based approach that learns to extract and combine an optimal set of image features for a given detection and labeling task.
 

Multi-parametric MRI Diagram

Automated tumor quantification. Multi-parametric data sources such as multi-contrast MRI (a) can be automatically fused by a machine learning based voxel classifier. When applied to novel data the result is a precise voxel labeling of the constituent tumor tissues that agrees well manual labelling by expert neuroanatomists (b). For tumor subtyping and survivability estimation additional information can be utilized such as diffusion MRI for cell density maps and white matter tract delineation and perfusion MRI for blood flow maps (c).
 

When applied to the task of labeling the components of brain tumors using a set of four magnetic resonance imaging contrast volumes (Figure a: T1, T2, T1 post Gd, and FLAIR) our approach achieves state-of-the art labeling accuracy (Figure b) and requires

To support future research, we derived additional maps (Figure c) that could be used to extend the approach to predict pathology subtypes and individual patient survival times.