Introduction/Background Multicenter prospective clinical validation of the radiomic machine learning model (TRACE4OC) applied to transvaginal ultrasound (US) in predicting the risk of malignancy of adnexal masses.

Methodology From a multicenter prospective consecutive series of women scheduled for surgery of adnexal masses, we collected and evaluated, fully blinded, 230 preoperative US images of adnexal masses with the TRACE4OC radiomic model previously developed according to the International Biomarker Standardization Initiative guidelines, trained and externally validated on a retrospective study of 274 US images of adnexal masses using histopathology as reference standard. Figure 1 shows the distribution of a radiomic texture feature (entropy of the co-occurrence matrix of gray levels) in an ovarian cystic malignant mass (a mucinous borderline tumor).

Results TRACE4OC model showed 91.3% accuracy, 99.0% sensitivity, 86.4% specificity when tested on the prospective multicentric external datasets of 230 masses (resulting into 90 malignant and 140 benign lesions at final histology), achieving 82.4% positive predictive value (PPV). The model shows a high correlation with finali histology (Pearson r: 0.8425 (95%CI: 0.800–0.876);p<0.001). The discrepancy was 0.473 ((SD: 0.50) 95%CI: 0.408, 0.538).

• Download figure
• Open in new tab
• Download powerpoint

Abstract 2022-RA-610-ESGO Figure 1

Conclusion The radiomic machine learning model can support clinicians in the diagnostic process of benignancy versus malignancy for adnexal masses, providing a strong reduction of the definite surgery rate for benign lesions still warranting very high sensitivity.