Doppler-based Radiomics Integrated With Frozen Section Analysis Enhances Borderline Ovarian Tumor Prediction: A Multicenter Study.

OBJECTIVE

To develop a Doppler-based radiomics model that utilizes tumor characteristics related to neovascularization, with or without integrating intraoperative frozen section analysis (FSA), to enhance borderline ovarian tumors (BOT) prediction.

METHODS

In this retrospective multicenter study, patients diagnosed with BOT or epithelial ovarian cancer (EOC) between January 2010 and October 2022 across five centers were included. Data from Center I were randomly divided into training (80%) and validation (20%) sets, while Centers II-V provided the test set. Three ultrasound-radiomics models were developed: color Doppler features (CD), grayscale features (GS), and combining both (GSCD). An Integration Model was developed by incorporating clinical variables, FSA and the best-performing radiomics model. Color Doppler signals were acquired through three channels with RGB information. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC).

RESULTS

A total of 1244 patients with a median age of 49 years (IQR, 37-58 years) were included, consisting of 520 patients with BOT and 724 EOC. The GSCD exhibited the best predictive performance for BOT among the three ultrasound-radiomics models, showing an AUC of 0.975 (95%CI, 0.962-0.985) in the training set, with a sensitivity of 92.9% and a specificity of 88.9% in the validation set. The Integration Model further improved the performance, achieving an AUC of 0.997 (95%CI, 0.993-0.998) in the training set, with a sensitivity of 87.2% and a specificity of 98.7% in the validation set. Additionally, both the GSCD and Integration Model effectively distinguished BOT from Stage I EOC, with AUCs of 0.919 (95% CI: 0.849-0.963) and 0.936 (95% CI: 0.871-0.974) in the validation set.

CONCLUSION

We developed a Doppler-based radiomics model utilizing red-green-blue information. This model, irrespective of integration with FSA, demonstrated improved accuracy in predicting BOT and stratified the FIGO Stage EOC.