This work was the third of three aims completed during my PhD.
Background: Tumors are highly heterogeneous, which can impact the effectiveness of delivered treatment. Quantitative imaging has shown promise for better characterizing tumor heterogeneity compared to qualitative, standard of care imaging; but existing efforts are either too complex for deployability or do not sufficiently capture heterogeneity.
Objective: We aimed to 1) identify distinct tumor habitats from multiparametric quantitative MRI using unsupervised clustering and 2) subsequently develop a simple set of ordinary differential equation (ODE) models to describe the growth of these identified habitats with and without radiation therapy.
Methods: We analyzed time series data from N=21 Wistar rats innoculated with C6 glioma, 8 of which were in the control group (no treatment). From diffusion-weighted and dynamic contrast-enhanced MRI data, we obtained four quantitative parameters, so each pixel in the segmented tumors at all time points had an associated 4-dimensional feature vector. Using MATLAB, we applied k-means clustering on the quantitative feature space to identify distinct tumor regions termed habitats. Once habitats were identified, we developed a family of ODE models to mathematically describe the habitat growth. We then used the Akaike Information Criterion to select the lowest error model and to subsequently evaluate the predictive capability of that model using two separate approaches. Please see the publication below for all details of the methods.
Results: From unsupervised clustering, we identified three significantly different (p<0.05) tumor habitats, named as follows: high vascularity-high cellularity (HV-HC), low vascularity-high cellularity (LV-HC), and low vascularity-low cellularity (LV-LC). From model selection, we arrived at a five parameter model that enabled us to predict tumor habitat growth.
Objective: We aimed to 1) identify distinct tumor habitats from multiparametric quantitative MRI using unsupervised clustering and 2) subsequently develop a simple set of ordinary differential equation (ODE) models to describe the growth of these identified habitats with and without radiation therapy.
Methods: We analyzed time series data from N=21 Wistar rats innoculated with C6 glioma, 8 of which were in the control group (no treatment). From diffusion-weighted and dynamic contrast-enhanced MRI data, we obtained four quantitative parameters, so each pixel in the segmented tumors at all time points had an associated 4-dimensional feature vector. Using MATLAB, we applied k-means clustering on the quantitative feature space to identify distinct tumor regions termed habitats. Once habitats were identified, we developed a family of ODE models to mathematically describe the habitat growth. We then used the Akaike Information Criterion to select the lowest error model and to subsequently evaluate the predictive capability of that model using two separate approaches. Please see the publication below for all details of the methods.
Results: From unsupervised clustering, we identified three significantly different (p<0.05) tumor habitats, named as follows: high vascularity-high cellularity (HV-HC), low vascularity-high cellularity (LV-HC), and low vascularity-low cellularity (LV-LC). From model selection, we arrived at a five parameter model that enabled us to predict tumor habitat growth.
