Discrimination of lipoma from atypical lipomatous tumor/well-differentiated liposarcoma using magnetic resonance imaging radiomics combined with machine learning

Purpose To evaluate the diagnostic capability of radiomics in distinguishing lipoma and Atypic Lipomatous Tumors/Well-Differentiated Liposarcomas (ALT/WDL) with Magnetic Resonance Imaging (MRI). Materials and methods Patients with a histopathologic diagnosis of lipoma (n = 45) and ALT/WDL (n = 20), who had undergone pre-surgery or pre-biopsy MRI, were enrolled. The MDM2 amplification was accepted as gold-standard test. The T1-weighted turbo spin echo images were used for radiomics analysis. Utility of a predefined standardized imaging protocol and a single type of 1.5 T scanner were sought as inclusion criteria. Radiomics parameters that show a certain level of reproducibility were included in the study and supplied to Support Vector Machine (SVM) as a machine learning method. Results No significant difference was found in terms of gender, location and age between the lipoma and ALT/WDL groups. Sixty-five parameters were accepted as reproducible. Fifty-seven parameters were able to distinguish the two groups significantly (AUC range 0.564-0.902). Diagnostic performance of the SVM was one of the highest among literature findings: sensitivity = 96.8% (95% CI 94.03-98.39%), specificity = 93.72% (95% CI 86.36-97.73%) and AUC = 0.987 (95% CI 0.972-0.999). Conclusion Although radiomics has been proven to be useful in previous literature regarding discrimination of lipomas and ALT/WDLs, we found that its accuracy could further be improved with utility of standardized hardware, imaging protocols and incorporation of machine learning methods.

Automated summary

The purpose of this study was to evaluate the diagnostic capability of radiomics in distinguishing lipoma and Atypic Lipomatous Tumors/Well-Differentiated Liposarcomas (ALT/WDL) with Magnetic Resonance Imaging (MRI). The results showed that the accuracy of radiomics could be improved by using standardized hardware, imaging protocols, and incorporation of machine learning methods.