Development and validation of a multiphase computed tomography-based radiomics classifier for differentiating...

We aimed to develop and internally validate a radiomics classification model based on multiphase computed tomography (CT) scans for preoperative differentiation of retroperitoneal non-fatty dedifferentiated liposarcoma (DDL) from leiomyosarcoma (LMS).

This retrospective study enrolled 78 DDL patients and 51 LMS patients who underwent surgical resection and pathological confirmation at our hospital between January 2011 and April 2023. Enhanced CT scans were performed within two weeks prior to surgery. An experienced radiologist manually delineated the tumor regions of interest using ITK-SNAP software on arterial-phase CT images, with contours copied to plain and venous-phase images. The dataset was split hierarchically (80% for training, 20% for testing). Radiomics features were extracted from non-contrast, arterial, and venous phases using pyradiomics 3.1.0 in Python 3.9.0, yielding 1130 features. LASSO regression was employed for feature selection with 5-fold cross-validation, retaining ≤ 10 features per imaging phase to prevent overfitting. Four machine learning models-logistic regression (LR), support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost)-were developed using scikit-learn 1.2.0. Model performance was evaluated based on the area under the curve (AUC), accuracy, sensitivity, and specificity. Statistical comparisons of AUC values between models, as well as between multiphase and single-phase models, were conducted using the DeLong test.

The LR and SVM models demonstrated superior performance across all evaluation stages, particularly when integrated with multiphase CT features. The multiphase SVM model achieved the best performance, with an AUC of 0.857, sensitivity of 0.812, specificity of 0.902, and accuracy of 0.847. DeLong tests further confirmed that the multiphase SVM model significantly outperformed its single-phase counterparts (non-contrast, arterial, and venous phases; all P < 0.05), while the multiphase LR model also exhibited significantly higher AUC than its non-contrast and venous-phase versions (both P < 0.05). Moreover, among all multiphase models, SVM showed significantly better discriminative ability than RF and XGBoost (both P < 0.05). These results validate that the comprehensive model integrating non-contrast and enhanced-phase CT features enhances both accuracy and reliability, outperforming single-phase models in distinguishing non-fatty DDL from LMS, with multiphase integration proving particularly advantageous for SVM and LR models.

Our CT radiomics-based classification model demonstrates promising potential for predicting the histological types of retroperitoneal non-fatty DDL and LMS. By integrating radiomics features from non-contrast and enhanced-phase CT images, this preliminary study suggests that such an approach could evolve into a valuable non-invasive diagnostic aid. However, further validation in larger, multicenter cohorts is essential to confirm its generalizability and clinical utility before routine application.