Radiomics of multi-parametric MRI for the prediction of lung metastasis in soft-tissue sarcoma: a feasibility study

Abstract

Purpose: To investigate the value of multi-parametric MRI-based radiomics for preoperative prediction of lung metastases from soft tissue sarcoma (STS).

Methods: In total, 122 patients with clinicopathologically confirmed STS who underwent pretreatment T1-weighted contrast-enhanced (T1-CE) and T2-weighted fat-suppressed (T2FS) MRI scans were enrolled between Jul. 2017 and Mar. 2021. Radiomics signatures were established by calculating and selecting radiomics features from the two sequences. Clinical independent predictors were evaluated by statistical analysis. The radiomics nomogram was constructed from margin and radiomics features by multivariable logistic regression. Finally, the study used receiver operating characteristic (ROC) and calibration curves to evaluate performance of radiomics models. Decision curve analyses (DCA) were performed to evaluate clinical usefulness of the models.

Results: The margin was considered as an independent predictor (p < 0.05). A total of 4 MRI features were selected and used to develop the radiomics signature. By incorporating the margin and radiomics signature, the developed nomogram showed the best prediction performance in the training (AUCs, margin vs. radiomics signature vs. nomogram, 0.609 vs. 0.909 vs. 0.910) and validation (AUCs, margin vs. radiomics signature vs. nomogram, 0.666 vs. 0.841 vs. 0.894) cohorts. DCA indicated potential usefulness of the nomogram model.

Conclusions: This feasibility study evaluated predictive values of multi-parametric MRI for the prediction of lung metastasis, and proposed a nomogram model to potentially facilitate the individualized treatment decision-making for STSs.

Keywords: Lung metastasis; MRI; Nomogram; Soft-tissue sarcoma.

Fig. 1

Feature selections from the multi-parametric MRI with LASSO. (A) Tuning parameter lambda selection in LASSO with 10-fold cross-validation. Vertical lines were drawn at the optimal values using the minimum criteria (left: the min criteria) and the 1 standard error of the minimum criteria (right: the 1-SE criteria). A λ value of 0.130, with log (λ), -2.040 was chosen (1-SE criteria). (B) LASSO coefficient profiles of the radiomics features, with 1-SE non-zero coefficients obtained from the MRI image. The vertical line was drawn at the optimal λ value which resulted in 5 non-zero coefficients

Fig. 2

Bar charts of the radiomics signature for each patient in the training (A) and validation (B) cohorts. Red bars indicated STS patients in the non-lung metastasis group, and green bars indicated STS patients in the lung metastasis group

Fig. 3

Construction and validation of the nomogram model. (a) The developed nomogram. (b) and (c) Calibration curves of the nomogram in the training (b) and validation (c) cohorts

Fig. 4

ROC curves of the margin, radiomics signature and nomogram in the training (a) and validation (b) cohorts

Fig. 5

DCA curves of the margin, radiomics signature and nomogram. The x-axis represented the threshold probability, whereas the y-axis measured the net benefit for the patients. The black line represented the hypothesis that all patients were without lung metastasis. The gray line indicated the hypothesis that all patients were with lung metastasis. The red line represented the nomogram. The blue line represented the radiomics signature. The green line represented the margin