Presented by: Bane Octavia MD
Icahn school of medicine, Mount Sinai

Introduction

While radiomics analysis based on magnetic resonance imaging (MRI) sequences has shown promise for small renal mass (SRM) characterization, clinical translation is challenging as radiomic feature stability, effect of observer and software package, and model performance remains uncertain. This has motivated our interest to evaluate radiomic features’ stability using inter-observer and inter-package measurement analysis in SRMs based on clinical MRI sequences and to employ machine learning (ML) statistical models for tumor characterization.

Materials

Thirty-one patients (21M/10F; age 60.9±10.4y) with SRMs (18 clear cell renal cell carcinoma (ccRCC)/6 non-clear cell (non-cc)RCC/7 benign; mean size, 3.2±1.8cm) undergoing surgery were recruited in a single-center prospective study. SRM volume-of-interest (VOI) regions were manually segmented on T2-WI, DWI/ADC, and T1-WI pre-/post-contrast imaging at 1 and 3-minutes using two separate radiomics software packages. Inter-observer measurements were obtained in a subset of 26 patients. A 3^rd observer performed a qualitative assessment including clear cell likelihood score (ccLS). Intra-class correlation coefficients (ICC) were calculated to assess inter-observer and inter-package reproducibility of radiomics measurements. Random forest models of radiomics and qualitative features were employed to distinguish RCC from benign SRM and ccRCC from non-ccRCC.

Results

,

Inter-observer comparison of radiomics measurements found that T1-WI pre/post-contrast and T2-WI yielded the greatest proportion of features with good/moderate ICC, while ADC measurements yielded low/moderate ICC. Inter-package comparisons demonstrated that most features had moderate/poor ICC, with the greatest stability found for measurements extracted from T1-WI post-contrast. ML radiomics models generated validation set AUCs range of 0.54-0.69 for RCC vs. other SRM and range 0.66-0.70 for ccRCC vs. non-ccRCC.

Conclusion

Radiomics features extracted from T1-WI pre/post-contrast demonstrates greater stability across observers and software packages, with fair/good accuracy in distinguishing RCC from benign SRM and ccRCC from non-ccRCC. Careful consideration to study design and methodology is required when conducting radiomics studies in MRI; data must be interpreted with caution when comparing radiomics analysis results using different software packages.

Funding

none