Assessment of Radiomics Feature Sensitivity to Field of View Variations In MRI Phantom Data
Abstract
Purpose
Radiomics features are increasingly used for quantitative image analysis; however, their robustness to image acquisition and reconstruction parameters remains a critical concern. Field of view (FOV) variations can alter image sampling and intensity distributions, potentially impacting radiomics feature values. This study investigates the variability in radiomics features to changes in FOV with the goal to use robust features to assess disease progression during and post HDR gynecological treatments.
Methods
A set of radiomics features was extracted from MR images of ACR MRI phantom. MRI images were reconstructed with four different FOVs (230, 300, 400, and 500). For each feature, values corresponding to the four FOVs were analyzed to assess variability and monotonic trends. Since only a single measurement per FOV was available, descriptive and non-parametric statistical metrics were employed. These included the mean, standard deviation, coefficient of variation (CV), percent change between FOV 230 and FOV 500, and Spearman rank correlation between feature values and FOV. Feature variation was visualized using line plots and logarithmic scaling to accommodate the wide dynamic range inherent to radiomics features.
Results
Substantial variability in feature sensitivity to FOV changes was observed. Several features exhibited minimal variation across FOVs, characterized by low coefficients of variation and near-zero percent change, indicating robustness to FOV selection. In contrast, other features demonstrated pronounced monotonic trends with increasing FOV, reflected by strong positive or negative Spearman correlation coefficients. Logarithmic visualization further highlighted differences in relative sensitivity among features spanning multiple orders of magnitude.
Conclusion
Radiomics features are sensitive to changes in field of view. Features demonstrating low variability across FOVs may be preferable for modeling and longitudinal studies, while highly FOV-dependent features should be used with caution or standardized acquisition protocols. This analysis provides a practical framework for evaluating radiomics feature stability in the absence of repeated measurements.