Whole-tumor 3D volumetric MRI-based radiomics approach for distinguishing between benign and malignant soft tissue tumors.

OBJECTIVES

Our purpose was to differentiate between malignant from benign soft tissue neoplasms using a combination of MRI-based radiomics metrics and machine learning.

METHODS

Our retrospective study identified 128 histologically diagnosed benign (n = 36) and malignant (n = 92) soft tissue lesions. 3D ROIs were manually drawn on 1 sequence of interest and co-registered to other sequences obtained during the same study. One thousand seven hundred eight radiomics features were extracted from each ROI. Univariate analyses with supportive ROC analyses were conducted to evaluate the discriminative power of predictive models constructed using Real Adaptive Boosting (Adaboost) and Random Forest (RF) machine learning approaches.

RESULTS

Univariate analyses demonstrated that 36.89% of individual radiomics varied significantly between benign and malignant lesions at the p ≤ 0.05 level. Adaboost and RF performed similarly well, with AUCs of 0.77 (95% CI 0.68-0.85) and 0.72 (95% CI 0.63-0.81), respectively, after 10-fold cross-validation. Restricting the machine learning models to only sequences extracted from T2FS and STIR sequences maintained comparable performance, with AUCs of 0.73 (95% CI 0.64-0.82) and 0.75 (95% CI 0.65-0.84), respectively.

CONCLUSION

Machine learning decision classifiers constructed from MRI-based radiomics features show promising ability to preoperatively discriminate between benign and malignant soft tissue masses. Our approach maintains applicability even when the dataset is restricted to T2FS and STIR fluid-sensitive sequences, which may bolster practicality in clinical application scenarios by eliminating the need for complex co-registrations for multisequence analysis.

KEY POINTS

• Predictive models constructed from MRI-based radiomics data and machine learning-augmented approaches yielded good discriminative power to correctly classify benign and malignant lesions on preoperative scans, with AUCs of 0.77 (95% CI 0.68-0.85) and 0.72 (95% CI 0.63-0.81) for Real Adaptive Boosting (Adaboost) and Random Forest (RF), respectively. • Restricting the models to only use metrics extracted from T2 fat-saturated (T2FS) and Short-Tau Inversion Recovery (STIR) sequences yielded similar performance, with AUCs of 0.73 (95% CI 0.64-0.82) and 0.75 (95% CI 0.65-0.84) for Adaboost and RF, respectively. • Radiomics-based machine learning decision classifiers constructed from multicentric data more closely mimic the real-world practice environment and warrant additional validation ahead of prospective implementation into clinical workflows.