Radiomic analysis of contrast-enhanced ultrasound data

Abstract

Radiomics describes the use radiological data in a quantitative manner to establish correlations in between imaging biomarkers and clinical outcomes to improve disease diagnosis, treatment monitoring and prediction of therapy responses. In this study, we evaluated whether a radiomic analysis on contrast-enhanced ultrasound (CEUS) data allows to automatically differentiate three xenograft mouse tumour models. Next to conventional imaging biomarker classes, i.e. intensity-based, textural, and wavelet-based features, we included biomarkers describing morphological and functional characteristics of the tumour vasculature. In total, 235 imaging biomarkers were extracted and evaluated. Dedicated feature selection allowed us to identify user-independent and stable imaging biomarkers for each imaging biomarker class. The selected radiomic signature, composed of median image intensity, energy of grey-level co-occurrence matrix, vessel network length, and run length nonuniformity of the grey-level run length matrix from the diagonal details, was used to train a linear support vector machine (SVM) to classify tumour phenotypes. The model was trained by using a four-fold cross-validation scheme and achieved 82.1% (95% CI [0.64 0.92]) correct classifications. In conclusion, our results show that a radiomic analysis can be successfully performed on CEUS data and may help to render ultrasound-based tumour imaging more accurate, reproducible and reliable.