Longitudinal DCE-MRI-based habitat imaging for early prediction of treatment response to neoadjuvant radiotherapy and targeted therapy in soft tissue sarcoma: a preliminary study

Early and reliable prediction of treatment response to neoadjuvant radiotherapy combined with targeted therapy remains a major clinical challenge in soft tissue sarcoma (STS). Conventional MRI and RECIST 1.1 have limited ability to detect early microvascular and microenvironmental alterations preceding measurable tumor shrinkage. This study evaluated whether longitudinal dynamic contrast-enhanced MRI (DCE-MRI)–based habitat imaging may serve as a non-invasive biomarker for early prediction of treatment response in STS. Twenty-eight patients with STS undergoing protocol-specified neoadjuvant radiotherapy-based treatment were prospectively enrolled. DCE-MRI was performed at baseline and 6 weeks post-treatment. Pathological response was assessed on resected surgical specimens by an experienced sarcoma pathologist according to the percentage of viable residual tumor cells. Pharmacokinetic parameters (Ktrans, Ve, Kep) were derived using the Tofts model. Tumor habitats were generated via K-means clustering of voxel-wise Ktrans and Ve maps. Radiomic features were extracted from whole-tumor masks and habitat subregions. Delta radiomics features were calculated as the difference between pre- and post-treatment feature values (pre-treatment minus post-treatment). Seven features associated with treatment response were selected through univariate analysis. Three logistic regression models (the Whole-tumor, Habitat, and Delta models) were constructed using AIC-based bidirectional selection, whereas RECIST 1.1 was evaluated separately as a fixed categorical clinical reference. Model performance was assessed using ROC curves, bootstrap internal validation, calibration curves (Hosmer–Lemeshow test), decision curve analysis (DCA), and exploratory DeLong tests. Favorable treatment response (≤ 5% viable tumor cells) was observed in 5 of 28 patients (17.9%). All habitats showed significant volumetric reduction following therapy (all P < 0.05), although proportional shrinkage did not differ among subregions ( P = 0.32). The habitat-based model, integrating pre-treatment Ktrans cluster-1 kurtosis and post-treatment Kep cluster-3 uniformity, demonstrated the highest apparent discriminative performance among the three radiomics-based models (AUC = 0.887; sensitivity = 100%; specificity = 73.9%). After bootstrap internal validation, the optimism-corrected AUC decreased to 0.764, and corrected discrimination was comparable across radiomics-based models. RECIST 1.1 yielded a sensitivity of 40.0%, specificity of 82.6%, and accuracy of 75.0% against the pathological response endpoint. The Habitat model also demonstrated favorable calibration (Hosmer–Lemeshow P = 0.973) and potential decision-analytic utility. In this prospective preliminary study, longitudinal DCE-MRI habitat imaging showed promising but exploratory performance as a candidate non-invasive biomarker for early pathological response assessment in STS. Validation in larger prospective multicenter cohorts with more balanced response groups and standardized treatment protocols is required. This study was based on two clinical trials prospectively registered on ClinicalTrials.gov ( https://clinicaltrials.gov/ ): NCT05167994 (2021-12-22) and NCT05938374 (2023-07-10).