Radiomics-based ultrasOund Model for differentiating Uterine Sarcomas from leiomyomas (ROMUS): a retrospective pilot...

Area of Science:

• Gynecologic Oncology
• Medical Imaging
• Machine Learning

Background:

• Distinguishing uterine sarcomas from benign leiomyomas preoperatively is challenging.
• Accurate differentiation is crucial for appropriate patient management and treatment planning.

Purpose of the Study:

• To develop and evaluate machine-learning models for differentiating uterine sarcomas from leiomyomas.
• Incorporate clinical data and radiomics features from ultrasound images.

Main Methods:

• Retrospective case-control study of 200 patients (100 uterine sarcomas, 100 leiomyomas).
• Extracted radiomics features (intensity, texture) and patient age.
• Developed and validated logistic regression, random forest, extreme gradient boosting (XGBoost), and support vector machine models.
• Compared model performance against subjective assessments by ultrasound examiners.

Main Results:

• Eight statistically significant, non-redundant radiomics features were selected.
• An XGBoost model integrating patient age and radiomics achieved an AUC of 0.93, sensitivity of 0.93, and specificity of 0.83 in the validation set.
• Model performance was comparable to subjective assessments by original and expert ultrasound examiners.

Conclusions:

• A machine learning model combining radiomics features and patient age shows good performance in differentiating uterine sarcomas from leiomyomas.
• The model's sensitivity is higher than specificity, similar to expert subjective assessment.
• Further prospective studies are needed to confirm radiomics' role and explore integration into clinical practice.