Evaluation of Diagnostic Accuracy of Preoperative CT-Based Radiomics in Primary Retroperitoneal Sarcoma.

Abstract

Obtaining an accurate preoperative diagnosis in retroperitoneal sarcoma (RPS) is challenging, and radiomics-based models may offer a valid option to achieve this outcome. This study evaluated the accuracy of radiomics-based preoperative CT models at predicting tumour histology and grade in patients with primary RPS. Data on consecutive patients who underwent surgery for primary retroperitoneal liposarcoma (RLPS) and leiomyosarcoma (RLMS) were analysed. Four different CT radiomics-based models were devised: 1) to distinguish between RLPS and RLMS; 2a) to predict overall tumour grade in both RLPS and RLMS; 2b and 2c) to predict tumour grade in RLPS and RMLS, respectively. The models were evaluated in a 100× random-split cross-validation. Data were available for 100 patients (64 RLPS, 36 RLMS), with 34 RLPS and 22 RLMS patients having a high-grade tumour on final histology. No significant differences in terms of age (p = 0.46), sex (p = 0.13), tumour location (p = 0.52), tumour diameter (p = 0.16), or tumour volume (p = 0.45) were observed between high- and low-grade tumours. The resulting area under the curve (AUC) at distinguishing between RLPS and RLMS was 0.94. The AUC at differentiating between high- and low-grade tumours for both RLPS and RMLS was 0.74. When tumour grade was analysed separately the corresponding AUC for RLPS and RLMS was 0.87 and 0.61, respectively. Radiomics-based preoperative CT-models were demonstrated to be accurate at differentiating between RLMS and RLPS, and at predicting preoperative tumour grade in RLPS, whereas they performed poorly in RLMS.