1. Planning. A radiology researcher can’t begin building a useful radiomic model without first naming a clinical problem in need of a solution, Elmahdy and Sebro state. More:

The study must have the correct study design and sufficient statistical power to appropriately evaluate radiomic models. A detailed literature review should be performed to ensure that the authors are aware of all previous work on the topic.”

2. Data curation. Optimally data used in radiomics research is diverse enough to represent the patient population at hand. The data should also be sourced from a variety of scanners and facilities, the authors maintain.

This helps create models that are able to be generalized to other similar patient populations and imaging devices. Data should, if possible, be manually reviewed to ensure no artefacts, incorrect data or incorrect labelling.”

3. Data pre-processing and image segmentation. Image pre-processing is best performed in a standardized way, Elmahdy and Sebro believe. Segmentation is amenable to manual work by a radiologist or other imaging expert, they add, or to automated or semi-automated approaches.

Regions of interest are used to define the specific region in which radiomic features are calculated. For segmentation analysis, DICOM images can be imported into several [off-the-shelf] programs. … Interpolation or estimation of the number of data points within the range of the dataset is important to reduce directional biases and ensure isotropic voxel spacing.”

4. Radiomic feature extraction. Ideally, radiomic features are reproducible such that the same feature values will obtain should the dataset be re-analyzed—evidently a not-uncommon scenario.

There are several thousand radiomic features that can be extracted from a single image. Feature reduction selects only the important radiomic features to be used in the final model and is therefore an important aspect of model building.”

5. Radiomic model building and validation. Internal datasets are best randomly split into training, validation and internal test data sets, the authors write. These should be separated from datasets used for external validation at the patient level to make sure internal and external sets are truly distinct from one another.

Model validation occurs after model training. Tuning is done as part of the validation step and is essential because poor tuning of the parameters may result in the model overfitting the data.”

6. Internal test. The optimal tuned model performance should be assessed in the internal test and external validation datasets, Elmahdy and Sebro recommend.

The area under the curve (AUC) and the F1 measure should be reported especially in cases of class imbalance. If the results are satisfactory the radiomic model should undergo external validation to assess generalizability.”

7. External validation. The optimal tuned model performance should also be evaluated in an independent external validation dataset, the authors note.

This step is essential to ensure the radiomics model generalizes to datasets other than the internal data set and that the model is not overfit to the training data set.”