Well-differentiated liposarcomas (WDLPS) can be difficult to distinguish from lipomas. Currently, this distinction is made by testing for MDM2-amplification, which requires a biopsy. The aim of this study was to non-invasively predict the MDM2-amplification status using radiomics, i.e. a combination of quantitative imaging features and machine learning techniques, thereby differentiating between lipomas and WDLPS.
Patients with a lipoma or WDLPS, a known MDM2-amplification status and a pre-treatment T1-weighted MRI scan who were referred to or diagnosed at the Erasmus MC from 2009-2018 were included. When available, other sequences, e.g. T2-weighted (fat-saturated) MRI, were included in the radiomics analysis. Imaging features describing intensity, shape, orientation and texture were extracted from the tumor segmentations; age, gender, tumor depth and tumor localization were added as semantic features. Classification was performed using an ensemble of various machine learning approaches. Evaluation was performed through 100x random-split cross-validation.
We included 116 patients: 58 patients with lipoma and 58 patients with WDLPS. The dataset originated from 41 MRI-scanners, resulting in large heterogeneity in imaging hardware and acquisition protocols. The best performing radiomics approach to differentiate between WDLPS and lipomas, based on T1-weighted imaging only, resulted in a mean [95% confidence interval] AUC of 0.75 [0.65-0.86], accuracy of 0.68 [0.59-0.77], sensitivity of 0.63 [0.48-0.78] and specificity of 0.73 [0.59-0.87]. A model based on the combination of imaging and semantic featured showed only a minor and non-significant improvement in performance.
There is a significant relationship between a radiomics signature, consisting of a combination of quantitative MRI features, and the MDM2-amplification status. Although further optimization and validation is needed for the transition to clinical practice, radiomics has appeared to be a promising, non-invasive approach for the classification of WDLPS and lipomas.
Clinical trial identification
M.Vos and M.P.A. Starmans contributed equally to this study.
Legal entity responsible for the study
Stichting Coolsingel, Netherlands Organisation for Scientific Research, Dutch Cancer Society.
All authors have declared no conflicts of interest.