Selection of optimal radiomics features for tumor phenotype differentiation using stochastic tumor growth modeling

Introduction: Due to various genetic and microenvironmental factors, cancers often manifest large variations in tumor phenotypes, with distinct progression patterns and sensitivities to treatment. As an example, hypoxic tumoral regions can diminish the effectiveness of therapies and degrade patient outcomes. Identifying such phenotype variations in-vivo can be difficult. This motivates development of radiomics-based analyses of PET images, with the aim to robustly classify tumors using their radiomics signatures, and subsequently, to select more effective treatment regimens. As such, there is a persistent need to identify which radiomics features and combinations thereof are optimal for phenotype discrimination. Using tumor growth simulations, we aimed to identify standardized radiomics features that can reliably detect hypoxia and necrosis in PET images of small tumors (< 1 cm) under realistic noise conditions.

Methods: We used a previously proposed hybrid, stochastic mathematical model of tumor growth to generate simulated cross-sections of tumors in normal tissue. The model includes a grid for cells and blood vessels, as well as partial differential equation grids for simulation of nutrient diffusion. Starting from a single cancer cell, growth was simulated until the tumor reached ~1 cm in diameter (65-164 days of simulated biological time). Four distinct tumor phenotypes were generated by varying blood vessel densities and angiogenesis rates (Fig. 1A): solid hypoxic (SH), solid necrotic (SN), diffuse hypoxic (DH), and diffuse normoxic (DN). Five random realizations of each phenotype were produced.

The simulated sequences of cell grids were converted to images of expected FDG uptake, forward-projected into sinogram space with added Poisson noise, and reconstructed with resolution 2.4 mm FWHM (Fig. 1B). The relative noise magnitude was 10% measured over a uniform region (modeling liver in clinical FDG PET scans).

The challenge is to reliably distinguish phenotype SH from SN, as well as DH from DN, as they appear very visually-similar in noisy PET images. For each image, we computed 22 Haralick features (HFs) available in the standardized PyRadiomics framework, and measured t-scores between feature distributions for different phenotypes. In addition, we tested over 200 paired feature combinations to find the optimal pair of features for phenotype discrimination.

Results: In noisy images, the best HFs to discriminate between phenotypes SH and SN (i.e. to detect necrotic cores) were SumEntropy (t = 3.1), JointEntropy (t = 2.6), JointEnergy (t = 2.5), and Correlation (t = 2.4). SumEntropy becomes sensitive to necrosis with tumor diameter starting from 8 mm, or 3.5 resolution units (Fig. 1C). To discriminate between phenotypes DH and DN (i.e. to detect hypoxia), the best features were Contrast (t = 8.5) and DifferenceAverage (t = 6.9).

In noise-free images, SumEntropy (t = 5.5) and Correlation (t = 5.1) were best for SH/SN discrimination, and DifferenceVariance (t = 11.0) and DifferenceEntropy (t = 10.3) were best for DH/DN discrimination.

To differentiate lesions of all 4 phenotypes, the best pair of features was Contrast and SumEntropy (Fig. 1D). The accuracy of SH/SN (DN/DN) phenotype classification was 100% (94%) without noise and 80% (94%) with 10% noise.

Conclusions: SumEntropy and Contrast were shown to be the most sensitive and noise-resilient HFs for distinguishing modeled tumor phenotypes with necrosis and hypoxia. These features can accurately classify small simulated lesions that appear visually similar in noisy images. Thus, our results imply that radiomics analysis can be beneficial even in small lesions extending to 3.5-4 resolution units. The proposed simulation framework will enable new investigations of the interplay between the tumor environment, phenotype, image noise, resolution, and region of interest definitions. It may also guide the development of new features that are informative, robust and reliable.

• Download figure
• Open in new tab
• Download powerpoint