Soft-tissue sarcomas (STS) are rare malignancies, accounting for approximately 1% of all cancers. Despite their rarity in aggregate, they represent a clinically pressing problem: roughly 40% of patients with high-grade STS of the extremities experience disease recurrence, and overall survival (OS) rates remain comparably low even when local disease control is excellent. Standard treatment for high-risk extremity STS combines surgical resection with neoadjuvant or adjuvant radiotherapy (RT), achieving local progression-free survival (LPFS) rates as high as 94%. Yet distant progression-free survival (DPFS) remains poor, and the gap between excellent local control and inferior systemic outcomes underscores the critical need for better tools to identify which patients carry the highest risk for metastatic disease and death.
Current limitations of staging: Treatment decisions in STS are routinely based on tumor grading (typically using the FNCLCC system) and tumor size, combined with the AJCC staging system. While grading and staging provide a useful framework, they capture only a fraction of the biological heterogeneity within STS. STS encompasses over 150 distinct histological subtypes, each with different molecular drivers and clinical behavior, which means that a single staging system applied across all subtypes will inevitably leave many high-risk patients unidentified at the time of initial diagnosis.
Two imaging-based approaches under investigation: This study, published in Cancers in 2021 by Peeken et al. from the Technical University of Munich (TUM) and the University of Washington (UW), formally compared two imaging-based strategies for prognostic assessment. The first is "semantic imaging" - the qualitative, expert-guided evaluation of MRI characteristics by trained radiologists. The second is "radiomics" - the automated, algorithm-driven extraction of hundreds of quantitative imaging features from segmented tumor volumes. Both approaches aim to extract prognostic information from pre-treatment MRI scans that goes beyond what conventional staging captures.
The study used two independent, retrospective cohorts from the two institutions, with one serving as the training set and the other as an external test set, giving the results a degree of validation rigor uncommon in single-center radiomic studies. This multicenter external validation structure positions the findings as more than preliminary, though important limitations related to histological heterogeneity and cohort differences remain.
Two independent retrospective cohorts formed the study backbone. The University of Washington (UW) cohort served as the training set (n = 108 patients) and the Technical University of Munich (TUM) cohort served as the external testing set (n = 71 patients). All patients had high-grade STS of the extremities or trunk, had received radiotherapy (neoadjuvant, adjuvant, or definitive) with curative intent, and had pre-RT MRI scans available. Strict exclusion criteria were applied: low-grade tumors, incomplete imaging, prior RT, primary bone sarcomas, Ewing sarcomas, rhabdomyosarcomas, distant metastases at diagnosis (M1), and endoprosthesis-related MRI artifacts. OS was measured from initial pathologic diagnosis to death or censoring. The study followed TRIPOD reporting recommendations.
Image acquisition and segmentation: All patients had pre-RT MRI scans including fat-saturated T2-weighted sequences (T2FS) and contrast-enhanced T1-weighted fat-saturated sequences (T1FSGd). Tumor volumes of interest (VOI) were manually segmented using MIM software, Eclipse, iplan RT, and 3D Slicer. Peritumoral edema was excluded from the VOI. To assess operator-dependent segmentation variability, 21 randomly selected UW patients received three independent delineations by separate radiation oncologists. The mean Dice similarity coefficient (DSC) across these multiple delineations was 0.92 (range: 0.81-0.96), confirming high inter-operator consistency.
Radiomic feature extraction: Preprocessing was standardized according to the Imaging Biomarker Standardization Initiative (IBSI) guidelines using the pyradiomics package (Version 2.2, Python 3.7). Steps included N4ITK MRI bias field correction, image discretization with a fixed bin width of 10, intensity normalization (mean redistribution with standard deviation scaling of 100), and isotropic resampling to 1x1x1 mm voxels using B-spline interpolation. Features were extracted only from original (non-filtered) images to avoid reproducibility concerns with filter-based features. In total, 105 radiomic features per sequence were extracted, encompassing first-order features, shape features, and five texture feature families: Gray Level Co-occurrence Matrix (GLCM), gray level size zone matrix (GLSZM), gray level run length matrix (GLRLM), neighboring gray tone difference matrix (NGTDM), and gray level dependence matrix (GLDM), yielding 210 features across both sequences.
ComBat batch harmonization: To address technical batch effects arising from different MRI scanner models across the two institutions, nonparametric ComBat harmonization was applied. This method estimates additive and multiplicative batch effects using a maximum likelihood approach and adjusts feature distributions accordingly, improving cross-site reproducibility of radiomic features before modeling.
Three musculoskeletal radiologists (with 8, 7, and 3 years of experience) independently evaluated MRI examinations while blinded to clinical information and histological diagnosis. The semantic features assessed covered multiple dimensions of tumor characterization: anatomical region (chest/back, leg, foot, arm, hand, gluteal/pelvic region), localization (epifascial, subfascial, or both; intramuscular; intermuscular), tumor morphology (multinodular, mass-like, or with superficial expansion), and tumor margins (well-defined, locally infiltrating, or diffusely infiltrating). On T1FSGd images, radiologists additionally scored extent of contrast enhancement (less than 1/3, 1/3-2/3, or greater than 2/3 of tumor volume), enhancement pattern (homogeneous or inhomogeneous), presence of vascularization, necrosis, perilesional contrast enhancement, and the "tail sign" (a well-defined, pointed curvilinear formation at least 10 mm in length on T1FSGd). On T2FS images, they assessed perilesional edema presence, edema diameter and extent, dominant T2FS signal intensity (hypointense/isointense/hyperintense), and signal pattern homogeneity.
Machine learning strategies compared: Three established survival analysis techniques were compared: elastic net regression (ENR), least absolute shrinkage and selection operator (LASSO), and random survival forest (RSF). All three were developed within a shared pipeline that combined feature reduction followed by model training. Feature reduction used 1000 bootstrap samples and applied a Spearman correlation threshold of 0.80 to exclude features correlated with AJCC staging and remove highly inter-correlated features. The Boruta algorithm was then applied to select the most relevant features. The final feature count per model was set as the median number selected across all bootstrap runs, yielding 12 features (Radiomics-T1), 10 features (Radiomics-T2), 11 features (Radiomics-T1T2), and 9 features (Semantic).
Cross-validation and hyperparameter optimization: Model comparison used 50 iterations of 5-fold nested cross-validation on the training cohort via the mlr3 R package. Hyperparameters were tuned using random search with 25 evaluations per model. For RSF, 1000 trees were used with optimization of node size (3-20) and mtry (2-10). For ENR, both alpha (0.05-1.0) and lambda were optimized. For LASSO, alpha was fixed at 1 and lambda was optimized. After identifying ENR as the best-performing strategy (mean rank 1.4 vs. RSF 2.2 and LASSO 2.4), final ENR models were retrained on the full training cohort using 5-fold cross-validation and evaluated on the independent TUM test set. The concordance index (C-index) was the primary performance metric, with 95% confidence intervals from 1000-fold bootstrapping.
Five model types were developed and compared: Radiomics-T1 (T1FSGd features only), Radiomics-T2 (T2FS features only), Radiomics-T1T2 (combined T1FSGd and T2FS features), Semantic (radiologist-assessed features only), and Radiomics-T1T2+Semantic (all imaging features combined). Combined imaging-clinical models additionally incorporated AJCC stage and age as covariates in multivariate Cox regression.
The two cohorts differed significantly in histological composition (p < 0.001), which represents an important source of model generalization challenge. Pleomorphic sarcoma was the dominant histology in both groups but more prevalent in the training set (45% vs. 34% in testing). In the training cohort, the next most common subtypes were leiomyosarcoma (11%) and spindle cell carcinoma (10%). In the testing cohort, myxofibrosarcoma (18%) and synovial sarcoma (13%) were more prominent. The testing cohort also carried more unfavorable clinical features: a larger proportion of AJCC stage 3 patients, 5 patients (7%) treated in a recurrent setting, a higher RT dose, and more patients receiving definitive RT (6% vs. 1%). The training cohort more frequently received chemotherapy. These imbalances represent a real-world scenario where cross-institutional models must generalize across meaningfully different patient populations.
Interrater agreement for semantic features: Ten patients from the TUM cohort were read by all three radiologists to assess reproducibility. For nominal and ordinal features, the median Fleiss Kappa was 0.524 (range: 0.035-1.00). Five features reached "substantial/good" agreement (Kappa > 0.60), five reached "moderate" agreement (0.40-0.60), and three reached only "slight/fair" agreement (Kappa 0.00-0.40). The two continuous measures (tumor diameter and edema diameter) had a median ICC of 0.833 (range: 0.138-0.846). This mixed agreement profile means that several semantic features used in model training may introduce systematic noise due to interrater variability.
Cross-validation performance (training set): Within the training cohort's nested cross-validation, Radiomics-T1 achieved a C-index of 0.68, outperforming Radiomics-T2 (C-index: 0.60). The Semantic model performed comparably to Radiomics-T1 in the training set (C-index: 0.67), suggesting that experienced radiologists can extract prognostically relevant information from MRI with similar discriminative power to automated radiomic features, at least within the training data distribution.
Feature selection findings: In the final Radiomics-T1 model, only two features retained non-zero coefficients: "Firstorder-Mean" (average voxel intensity) and "Shape-SurfaceArea" (the 3D surface area of the tumor volume). The Radiomics-T2 model selected several features from the GLSZM and GLDM families, which capture the spatial distribution and clustering of gray-level values across the tumor volume, reflecting textural heterogeneity.
When models were applied to the independent TUM test set, the performance ordering changed substantially compared to training. Both radiomic MRI models performed similarly: Radiomics-T1 achieved a C-index of 0.64 and Radiomics-T2 achieved a C-index of 0.63. These results exceeded both the Semantic model (C-index: 0.58) and the AJCC staging system alone (C-index: 0.61), demonstrating that radiomic features generalize better than expert-derived semantic features when applied across institutions with different patient populations and imaging protocols. Combining both MRI sequences (Radiomics-T1T2) did not improve performance further (C-index: 0.60), suggesting diminishing returns from combining feature sets rather than a consistent additive benefit.
Semantic model performance drop: The Semantic model failed to reproduce its training-set performance (C-index 0.67 training vs. 0.58 testing), a drop of 0.09 C-index points. The combined Radiomics-T1T2+Semantic model also did not improve beyond the individual radiomic models (C-index: 0.60). This generalization failure for the semantic model is consistent with known challenges in qualitative radiological assessment: moderate interrater agreement within the training cohort does not guarantee consistent feature application in an external cohort where different readers, different scanners, and different histological distributions interact.
Clinical baseline comparison: Among standalone predictors, age achieved the highest C-index in the test set (0.69), exceeding all imaging models. Tumor volume (C-index: 0.59) performed slightly worse than the AJCC staging system. Kaplan-Meier survival analysis using the median predictor value from the training set as the cutoff showed significant patient stratification for Radiomics-T1, Radiomics-T2, and Radiomics-T1T2 in the testing cohort (all log-rank p < 0.05). The Semantic model, Radiomics-T1T2+Semantic, Volume, and Age did not achieve statistically significant Kaplan-Meier separation, while AJCC showed only a trend (p = 0.0532).
Significant semantic features in univariate analysis: Three individual semantic features showed statistically significant associations with OS in univariate Cox proportional hazards regression on the combined cohort: "maximal diameter without tail" (p = 0.022), "necrosis" (p = 0.039), and "perilesional edema" (p = 0.043). None of these associations reached significance after Bonferroni correction for multiple testing, but they informed the feature selection pipeline. The "tail sign" - a radiological indicator of tumor infiltration along the deep fascia - was included in the final semantic feature set based on its prognostic relevance in prior literature on myxofibrosarcoma and undifferentiated sarcomas.
While imaging-only models provide moderate prognostic discrimination, the greatest benefit came from combining radiomic features with established clinical variables. The AJCC staging system combined with age (AJCC+Age) already achieved a C-index of 0.71 in the test set, making it the strongest model without imaging features. Adding the Radiomics-T2 model to AJCC+Age improved C-index by a further +0.02 to 0.73 (Radiomics-T2+AJCC+Age). This combined model also achieved the best mean time-dependent AUC of 0.79, and demonstrated statistically significant Kaplan-Meier patient stratification in the test set. Adding the Radiomics-T1 model to AJCC+Age failed to replicate this improvement (C-index: 0.67), and the Semantic+AJCC+Age combination performed even worse (C-index: 0.62), confirming that T2FS-based radiomics offers the most reliable incremental value when combined with standard clinical prognostic factors.
Calibration analysis: Model calibration was assessed using Brier scores and calibration curves. The clinical models (AJCC and Age) showed the lowest Brier scores (24 and 30, respectively) and the most monotonous calibration slopes. The radiomic models had Brier scores ranging from 88 to 114, while semantic feature models showed the worst calibration (Brier scores from 370 to 702). This poor calibration for semantic models means that while they can discriminate between risk groups at a group level, their absolute predicted risk values are poorly correlated with observed event frequencies. Combined imaging-clinical models substantially improved calibration (Brier scores from 37 to 56), underscoring the practical necessity of combining radiomic predictions with clinical variables rather than deploying imaging models in isolation.
Re-analysis with non-independent cohorts: To isolate the contribution of cross-site heterogeneity to model performance, the authors retrained and tested all models using randomly sampled, institution-mixed cohorts of equal size and event rates. Under this less stringent TRIPOD type II validation approach, the Semantic model's testing AUC improved to 0.63 (a difference of only -0.01 vs. training), and both radiomic models achieved testing AUCs of 0.63, matching their training performance. This substantial improvement in reproducibility for the Semantic model under mixed-cohort validation indicates that differing patient characteristics, histological distributions, and acquisition protocols between the original independent cohorts were a primary driver of the semantic model's generalization failure, not intrinsic instability in the semantic approach itself.
Retrospective design and selection bias: Both cohorts were collected retrospectively, introducing potential selection bias. To achieve clinically homogeneous patient groups meeting the strict inclusion and exclusion criteria (high-grade, extremity/trunk, pre-RT MRI available, no low-grade, no bone sarcoma, no Ewing, no rhabdomyosarcoma, no M1 disease), the overall patient numbers had to be reduced substantially. The training cohort of 108 patients required simultaneous availability of T1FSGd and T2FS sequences, which reduced the cohort size by 34% compared to a prior study using only T1FSGd. Smaller cohorts reduce statistical power and increase the risk that selected radiomic features may be dataset-specific rather than biologically generalizable.
Histological heterogeneity: STS is not a single disease but a collection of over 150 distinct histological subtypes, each with different biology, molecular drivers, and imaging characteristics. In this study, pleomorphic sarcoma dominated both cohorts, but the second and third most common subtypes differed substantially between training and testing sets. Building a histology-agnostic prognostic model across mixed subtypes is likely suboptimal, as the prognostic relevance of individual semantic or radiomic features almost certainly varies by subtype. The authors acknowledge that histology-specific models would be preferable, but single-center cohorts rarely contain sufficient cases of any individual rare STS subtype to support subtype-stratified modeling.
Technical heterogeneity and reader variability: The multicenter nature of the study introduced substantial technical heterogeneity across MRI scanner types and imaging protocols, which ComBat harmonization partially addressed. However, residual scanner-related feature variability may still influence radiomic predictions. For semantic features, reader variability (median Fleiss Kappa of 0.524, with three features showing only slight/fair agreement) adds noise to the semantic modeling pipeline. Additionally, the third radiologist read imaging from both cohorts, potentially introducing a degree of false cross-cohort consistency that artificially inflated semantic model reproducibility in the institutional comparison.
Survival endpoint and small event numbers: OS is the most clinically meaningful endpoint, but it requires longer follow-up and produces fewer events in smaller cohorts than intermediate endpoints like local recurrence. Small event numbers limit model complexity, reduce the reliability of hazard estimates, and increase variability in cross-validation performance metrics. Future STS prognostic studies should ideally be prospectively designed with pre-specified endpoints, well-defined acquisition protocols, and restriction to individual STS histologies to address these limitations.
Histology-specific radiomic models: The most important near-term direction identified by the authors is development of models restricted to individual STS histological subtypes rather than pan-histology approaches. Myxofibrosarcoma, liposarcoma, synovial sarcoma, and leiomyosarcoma each have distinct imaging characteristics and molecular drivers. Histology-specific models would capture subtype-relevant radiomic features more precisely and avoid the signal dilution that occurs when diverse tumor types are pooled. Realizing this goal requires multi-institutional database expansion, which the authors describe as an ongoing effort. Alternatively, histology could be incorporated as an explicit predictor variable in models trained on sufficiently large mixed-histology cohorts.
T2FS sequences as a contrast-free alternative: A clinically meaningful finding of this study is that T2FS-based radiomic models (which do not require gadolinium contrast agent administration) achieved predictive performance comparable to T1FSGd-based models in the test set (C-indices: 0.63 vs. 0.64). This is practically significant because contrast agent use carries risks including allergic reactions and, with repeated use, gadolinium deposition in brain tissues. The availability of a contrast-free radiomic prognostic tool would expand applicability to patients with contrast allergies, renal insufficiency, or other contraindications. Future prospective studies should specifically validate T2FS-based radiomic models in contrast-contraindicated patients.
Radiogenomics and delta radiomics: Beyond direct prognostic assessment, the authors outline two adjacent radiomic applications in STS. Radiogenomics refers to the prediction of molecular tumor characteristics (such as MDM2 amplification in liposarcoma or grading in mixed-histology cohorts) from imaging features without biopsy. The authors reference ongoing work demonstrating that MDM2 gene amplification status, which distinguishes atypical lipomatous tumors from benign lipomas, can be predicted from MRI-based radiomics. Delta radiomics, the analysis of longitudinal changes in radiomic features measured serially during or after RT or systemic therapy, offers a route to early response assessment. First studies in STS and osteosarcomas have shown promising early results for this approach.
Imaging-guided treatment intensification: The clinical end-goal articulated in this paper is imaging-guided treatment personalization. High-risk patients identified by radiomic biomarkers could be candidates for treatment intensification, such as addition of systemic chemotherapy, enrollment in clinical trials testing immunotherapy (the Sarc032 trial using Pembrolizumab, NCT03092323), or targeted molecular agents (MDM2 inhibitor AMG 232 in the NRG DT001 trial, NCT03217266; trabectedin in the TRASTS trial, NCT02275286). Low-risk patients could potentially be spared toxicities from treatment intensification. The authors caution that testing these radiomic-guided treatment selection strategies requires prospective trial designs with pre-specified imaging biomarker thresholds, and that the current retrospective evidence base is insufficient to support immediate clinical adoption.