Finding the Pieces to Treat the Whole: Using Radiomics to Identify Tumor Habitats

See also the article by Prior et al in this issue.

Hersh Sagreiya, MD, is an assistant professor of radiology at the University of Pennsylvania. He performs clinical work in abdominal imaging and research in machine learning and informatics, which involves developing novel deep learning algorithms, implementing artificial intelligence via PACS, and analyzing large-scale medical datasets for new linkages between imaging, genetics, and clinical data. He has received grants from RSNA, the Society of Radiologists in Ultrasound, and the American Lung Association.

Intratumoral heterogeneity refers to the areas within a tumor with distinct subpopulations of cells that exhibit characteristic genetic, epigenetic, and environmental influences. These areas can vary in their response to therapy, which can result in differing patterns of treatment resistance to targeted therapies, permit disease progression, and present a major therapeutic challenge (1). Habitat imaging seeks to identify these distinct tumor microenvironments using quantitative imaging techniques to provide insights into tumor phenotype and the interaction of the tumor with its microenvironment (2).

Identifying tumor habitats typically makes use of radiomic features, which involves generating quantitative data from medical images. However, repeatability and reproducibility, consistent measurements under the same and different conditions, respectively, are major challenges in the field of radiomics (3). The use of three-dimensional (3D) radiomic features has potential advantages over two-dimensional features, as 3D features can capture voxel relationships in the z-axis. For instance, a study of Ewing sarcoma found that 3D MR radiomic features had higher reproducibility and higher accuracy for predicting neoadjuvant chemotherapy response than two-dimensional features (4).

In a study in this issue of Radiology: Artificial Intelligence, Prior and colleagues analyzed 2436 hepatic and pulmonary lesions from 331 patients within four cancer cohorts: colorectal, lung, neuroendocrine, and a mixture of cancers (5). A radiologist segmented all measurable tumors as defined by Response Evaluation Criteria in Solid Tumors, version 1.1, and 91 features were analyzed. Repeatability was assessed by retesting the calculation of 3D radiomic features after performing image perturbations, including image rotation, translation, and noise addition. The authors set the kernel radius (R, which relates to the number of neighboring voxels) to either 1 mm or 3 mm and the bin size (B, which relates to the number of gray levels analyzed) to either 12 HU or 25 HU. This resulted in four scenarios: R1B12, R1B25, R3B12, and R3B25, with R in this nomenclature reflecting kernel radius and B reflecting bin size. Repeatability was assessed by comparing radiomic features before and after image perturbation (representing retesting) in all four scenarios. While the R3B12 combination had the highest repeatability, overall repeatability was poor. Reproducibility against bin size was assessed by holding kernel radius (R) constant and evaluating radiomic feature stability with varying bin size (B of 12 HU or 25 HU); reproducibility against kernel radius was assessed by holding bin size (B) constant and evaluating radiomic feature stability with varying kernel radius (R of 1 mm or 3 mm). Reproducibility against kernel radius was poor, but reproducibility against bin size was excellent. Twenty-six radiomic features with adequate repeatability and reproducibility against both bin size and kernel radius using the intraclass correlation (with lower confidence limit greater than or equal to 50%) were deemed precise. This resulted in a different set of precise features for liver and lung.

Unsupervised learning is a technique for analyzing patterns in unlabeled data. The authors used an unsupervised machine learning model based on the Gaussian mixture model to determine tumor habitats. They investigated specifically whether using precise radiomic features (those with sufficient repeatability and reproducibility) resulted in improved habitat formulation compared with using nonprecise features (all radiomic features without such filtration). Habitat stability was assessed by calculating the Dice similarity coefficient between habitats generated from the original and perturbed images. The CT habitats created using precise features were more stable for both liver and lung lesions. As an exploratory analysis, the CT habitats that the authors generated in an independent cohort of 13 patients demonstrated similarity with both multiparametric MRI habitats and digitized images using hematoxylin-eosin staining from biopsy.

Radiomic features computed using a larger kernel size were more repeatable and reproducible in this study; one reason may be that a larger kernel incorporates more pixels for the computational analysis and is hence able to capture more tumor heterogeneity in the surrounding microenvironment. As described in Appendix S3, a kernel radius of 1 mm generates a cubic kernel in the neighborhood of a given voxel of size 3 × 3 × 3, and a kernel radius of 3 mm generates a cubic kernel of size 7 × 7 × 7. That represents 12.7 times as much local data that are being analyzed due to the cubic relationship between radius and volume. In the same vein, a prior study involving the radiomic analysis of lung cancer on CT scans found that the reproducibility of voxelwise radiomic features in test-retest data, as measured by the concordance correlation coefficient, was more sensitive to changes in kernel radius than bin width (6).

Regarding bin size, features computed using the smaller bin size of 12 HU were more repeatable and reproducible than those with bin size of 25 HU, but the difference was far smaller than it was for kernel size. It may be that much of the underlying tumor heterogeneity is already captured with a bin size of 25 HU, so the relative improvement of going to 12 HU is minor. It would be interesting to see if such trends persist in tumor environments other than liver and lung, such as bone. For comparison, a study analyzing the stability of radiomic features on MRI apparent diffusion coefficient maps of cervical cancer found that when feature normalization was performed prior to feature extraction, a fixed bin width approach with smaller bin widths was the recommended technique (7). A future direction for this study could involve testing a larger array of values for kernel radius and bin size. For instance, one could imagine that there is some value for bin size after which the performance starts to degrade, and this could vary by tumor environment or imaging technology (CT, MRI, US, PET, etc).

The authors showed that preselecting for precise radiomic features resulted in more stable CT habitat generation. For another common clinical task, a prior study showed that radiomic features with high repeatability demonstrated greater generalizability for prognostic models developed for nasopharyngeal carcinoma (8). Overall, it makes sense that models with more robust features perform better in various clinical tasks. In contrast to other articles in the literature, most of which focused on a single tumor and a large number of which were on non–small cell lung cancer or oropharyngeal cancer, this study examined multiple primary tumors in two distinct tumor environments, liver and lung, with notable differences in the most salient radiomic features in both of these environments, regardless of primary tumor (9). As the authors mention, this may very well be related to the higher degree of contrast between the tumor and background tissue for lung compared with liver. It would be valuable to see if this pattern persists for other tumors upon a soft tissue background, such as kidney or pancreas.

The lack of significant differences between primary tumors suggests that their metastases may have a similar appearance on CT scans. A future direction could involve a radiomic analysis distinguishing between hypovascular and hypervascular metastases, which more clearly vary in their appearance. Finally, the case series comparing habitats generated from CT and MRI with histology demonstrated intriguing similarities, although further research on a larger sample size is warranted. Noting the challenge involved in assembling such a dataset, a multi-institutional dataset of hundreds (or even thousands) of patients with habitats generated from CT and multiparametric MRI paired with histology could go a long way to establish the clinical utility of this approach and serve as a benchmark for future studies.

Overall, Prior and colleagues offer important insights regarding the factors that most affect radiomic feature repeatability and reproducibility, crucial for developing stable habitats that capture tumor heterogeneity, while serving as a springboard into future work on this important topic.