E427: Computed Tomography-Based Radiomic Features for Prognostication in Patients with Primary Retroperitoneal Sarcoma

Individual prognostication of patients with primary retroperitoneal sarcoma (RPS) relies mainly on clinical and pathological variables combined in prognostic nomograms. The objectives of this study were (1) to investigate whether computed tomography (CT)-based radiomic features could enhance the performance of the Sarculator nomograms for patients with primary RPS and (2) to investigate whether radiomic features could identify patients with high risk RPS (leiomyosarcoma, LMS, or G3 dedifferentiated liposarcoma, DDLPS).

Methods:

Patients with primary RPS treated with curative-intent surgery (2011-2015) at a single institution and with the preoperative CT scan available were included. Preoperative non-contrast enhanced CT scans were reviewed and segmented for extraction of 89 radiomic features. The selection of radiomic features was based on outcome-specific random forest models, through generation of multiple contexts where patients were split in a training and a testing set and including the nomogram Sarculator with a high selection probability (Figure 1). The performance of the prognostic models was assessed in terms of discrimination only (Harrell C index) and both discrimination and calibration (IPA).

Results:

The study population included 112 patients: 50% DDLPS, 20% well differentiated LPS, 20% LMS. Median follow-up was 77 months (IQR 65-92 mo). When the Sarculator nomograms were applied to the study cohort, IPA was 0.142 for DFS and 0.217 for OS.

The radiomic feature original_GLCM_MCC, a measure of texture complexity, enhanced the prediction accuracy of Sarculator for DFS (IPA =0.225; IPA gain: 0.083).  Although radiomic features could predict patient OS, their addition to the Sarculator OS nomogram had minimal improvement on the discriminative ability (IPA =0.229; IPA gain: 0.012). We then tested if radiomic features could aid identification of high risk patients (G3 DDLPS and LMS). Among the radiomic features, original_firstorder_10Percentile, which represented the distribution of gray values in CT scan, ranked first to identify patients with high risk RPS. The accuracy of this feature was 0.85 and the negative predictive value was 0.88 when optimized to lower the probability of a false negative prediction.


Conclusions:

The integration of radiomic features marginally improved the performance of the Sarculator DFS nomogram for patients with primary RPS. Radiomic may help identifying high risk RPS in the preoperative setting, when the percutaneous biopsy might not be able to properly assess tumor grade.