Table 3.

Recommendations for conducting and reporting studies that investigate radiogenomic associations with tumor immune phenotypes.

Process	Considerations	Recommendations
Study design	Study registration	Pre-register studies in databases such as the Open Science Framework (OSF)
	Cohort selection	Focus on specific molecular subtypes or subclasses of cancers may enable more accurate radiogenomic modelsMeta-analysis of multiple cohorts can be used to achieve more generalizable models
	Study design	Prospective study design to enable longitudinal feature assessment may be ideal for generating models to predict immunotherapy response and identify biomarkers of resistanceFor retrospective study design, statistical and modeling approaches should be decided a priori
Evaluating molecular data	Tumor and TME gene expression data procurement and processing	RNA-seq for assessing gene expression, refer to Conesa et al. 2016 for a review of good data practices [121] RNA-seq may be eventually supplanted by single-cell RNA-seq, which can improve the ability to distinguish tumor versus immune cell gene expression
	Pathway and immune infiltration analysis	Software like Gene Set Enrichment Analysis (GSEA), Ingenuity Pathway Analysis, DAVID, Metascape are standard for pathway enrichment analysis Approaches including single sample GSEA (ssGSEA), CIBERSORT, and Immunoscore useful for more specific quantification of types of tumor immune cell infiltration
	Cell markers by IHC	Specific staining of cell surface markers remains the gold standard for quantifying immune cell infiltration To increase staining throughput, consider using tissue microarrays and multiplexed IHC
	Quantifying TILs by H&E	H&E allows for good quantitation of TILs, but is often subject to clinician-reader bias Best clinical practices are outlined in Salgado et al. 2015 [122]
Image acquisition, processing, and extraction	Image acquisition parameters	Use standardized acquisition parameters
	Image pre-processing	Normalize voxel intensities of images, particularly MRI, to more accurately and reproducibly extract radiomic features
	Feature definition and extraction	Use feature standardization platforms, such as MITK Phenotyping and the Image Biomarker Standardization Initiative
	Tumor segmentation	Use multiple independent observers if segmenting manually or consider semi-automatic/automatic approaches to maximize reproducibility
	Deep learning	Utilize algorithm visualization methodology, such as saliency maps, to increase interpretability/explainability/transparency
Modeling and data analysis	Feature selection	Reduce feature dimensionality such as through regression modeling (e.g. LASSO Cox, Elastic Net) or using intra-class feature similarity measures (e.g. intra-class correlation coeffcient) to prevent overfitting and improve feature reliability
	Model design	Best performing models for predicting prognosis and immunotherapy response are likely achieved by combining radiogenomics models with other covariates into composite models Correct for multiple hypothesis testing where appropriate
	Machine learning	Use hold-out data sets for evaluation of models and to prevent any data leakage from training to evaluation sets Validate on data that are independent from the training set and ideally from multi-institutional sources
Data transparency and reporting	Public data and code repositories	Share code in open-source repositories like GitHub Share imaging data in public repositories like the Imaging Data Commons (IDC)
	Radiomics quality score (RQS)	Report RQS score (out of 36) developed by Sanduleanu et al. 2018 [1]
	Study reporting checklists	Use of TRIPOD 22-item checklist for model development and validation [120]

Legend: DAVID: database for annotation, visualization, and integrated discovery, IHC: immunohistochemistry, TIL: tumor infiltrating lymphocyte, H&E: hematoxylin and eosin, MITK: medical imaging interaction toolkit, LASSO: least absolute shrinkage and selection operator, TRIPOD: Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis.