Soft tissue sarcomas are a rare and biologically heterogeneous group of malignancies arising from mesenchymal tissues. Their imaging appearance varies dramatically across subtypes, anatomic locations, and degrees of aggressiveness, making noninvasive characterization difficult. Initial evaluation typically begins with ultrasound to assess the mass and determine its size, depth, and margins, followed by radiography to evaluate for bone involvement. Suspicious lesions, particularly deep or large masses, proceed to contrast-enhanced MRI, which provides the most detailed characterization of tumor matrix and margin definition. CT is used for chest staging and occasionally for local staging of retroperitoneal tumors.
Limitations of conventional imaging: Traditional semantic descriptors of tumor morphology, including size, margin definition, and signal intensity on MRI, offer only a limited representation of the underlying tumor biology. A standard measure such as signal intensity standard deviation gives a global summary statistic that fails to account for the spatial arrangement of pixel intensities within the tumor. Two tumors with entirely different internal textures can have identical means and standard deviations if they contain the same numbers of high- and low-intensity voxels arranged differently. This is a fundamental reason why conventional imaging struggles to grade sarcomas, predict metastases, or assess treatment response quantitatively.
What radiomics adds: Radiomics is the subfield of medical imaging dedicated to extracting high-dimensional quantitative features from imaging data using computational algorithms. These features encompass descriptors of intensity distribution (first-order statistics), spatial relationships between pixel intensities (texture features derived from gray-level co-occurrence matrices, gray-level run-length matrices, and similar constructs), and geometric shape descriptors. Because they capture spatial heterogeneity that is invisible to the human eye, radiomic features can encode information about tumor cellularity, vascularity, necrosis, and microenvironment composition. Advanced imaging modalities such as diffusion-weighted imaging (DWI) and dynamic contrast-enhanced (DCE) MRI add additional dimensions but are constrained by motion artifacts, value overlap across tumor types, and the fact that they require dedicated acquisition protocols not available at all centers.
This 2025 review in Cancers, authored by Shah et al. from institutions including the University of Miami and St. Jude Children's Research Hospital, surveys the current state of radiomics applied to soft tissue sarcoma diagnosis, grading, prognostication, treatment planning, and treatment response assessment. It also addresses the reproducibility challenges and workflow integration barriers that separate research-stage radiomics from routine clinical use.
A radiomics pipeline begins with image acquisition, proceeds through tumor segmentation, advances to feature extraction, and culminates in model development and validation. Each step introduces potential variability that can affect the reproducibility and generalizability of the resulting model. The first step, segmentation, involves delineating the tumor boundary on each imaging slice to create a three-dimensional region of interest (ROI). Segmentation can be performed manually by a radiologist, semi-automatically with user initialization, or fully automatically with deep learning algorithms. Interobserver variability in manual segmentation is one of the most important sources of radiomic feature instability, with some first-order and texture features changing substantially depending on who drew the contour.
Feature extraction: Once the ROI is defined, feature extraction algorithms compute hundreds to thousands of quantitative descriptors. First-order statistics (mean, median, kurtosis, skewness, interquartile range) summarize the distribution of voxel intensities within the tumor. Texture features from gray-level co-occurrence matrices (GLCM) describe how often pairs of pixels with specific intensity values and spatial relationships occur within the tumor. Other feature families include gray-level run-length matrices, gray-level size zone matrices, and Laplacian of Gaussian-based features that capture edge-related heterogeneity at multiple spatial scales. Shape features describe tumor geometry: volume, surface area, compactness, elongation, and sphericity.
The Image Biomarker Standardization Initiative (IBSI): Because different software packages historically used varying definitions for the same radiomic feature names, the IBSI established consensus definitions and reference benchmarks so that features computed by different tools would be directly comparable. Adherence to IBSI standards is now increasingly expected in high-quality radiomics studies and is directly relevant to soft tissue sarcoma research, where datasets are small enough that any systematic feature computation error can meaningfully corrupt model training.
Model development and feature selection: After extraction, high-dimensional feature sets (often thousands of features for a dataset of dozens of patients) require aggressive dimensionality reduction to avoid overfitting. Filtering approaches such as intraclass correlation coefficient (ICC) thresholds remove features that vary significantly between segmentation attempts. Redundancy reduction via pairwise correlation analysis removes collinear features. Supervised feature selection algorithms such as least absolute shrinkage and selection operator (LASSO), minimum redundancy-maximum relevance (mRMR), and principal component analysis (PCA) then identify the subset of features most predictive of the clinical endpoint, balancing model complexity against training sample size.
A fundamental clinical application of radiomics in sarcoma is determining whether a soft tissue mass is benign or malignant before biopsy. Conventional qualitative imaging can suggest malignancy in certain subtypes based on specific features, such as the "tail sign" extending along fascial planes in myxofibrosarcoma or the "triple sign" (solid cellular elements plus hemorrhage or necrosis plus fibrotic regions) in synovial sarcoma. However, the majority of soft tissue sarcomas cannot be confidently subtyped by imaging appearance alone. A meta-analysis of 10 studies pooling 885 subjects found that MRI radiomics distinguishes benign from malignant soft tissue masses with sensitivity of 84%, specificity of 83%, and AUC of 0.93, a performance level that supports using radiomics to triage which masses require biopsy.
Uterine leiomyosarcoma versus leiomyoma: One of the most clinically impactful radiomics applications in sarcoma is distinguishing uterine leiomyosarcoma from benign leiomyoma (including atypical or degenerated leiomyomas). This distinction is critical because needle biopsy is often contraindicated in suspected leiomyosarcoma due to the risk of peritoneal tumor seeding. Even without radiomics, leiomyosarcomas display characteristic MRI features including lower apparent diffusion coefficient (ADC) values, more heterogeneous contrast enhancement, and increased signal intensity on T1-weighted images from internal hemorrhage, but these features overlap with atypical leiomyomas. A multiparametric MRI plus radiomics model by Roller et al. achieved AUC 0.989, modestly improving over a model using MRI and clinical data only (AUC 0.956). A CT-based machine learning study achieved AUC 0.78-0.97 with sensitivity up to 100% and specificity up to 93%, outperforming radiologists alone (AUC 0.73-0.75). A combined model for T2 hyperintense uterine mesenchymal tumors reached AUC 0.91, outperforming radiologists (AUC 0.78-0.90), clinical data alone (AUC 0.79), and radiomics alone (AUC 0.76).
Other sarcoma subtypes: The diagnostic differentiation benefit extends across multiple sarcoma types. Timbergen et al. demonstrated a radiomics model distinguishing desmoid-type fibromatosis from soft tissue sarcomas across 203 cases with AUC 0.88, comparable to two radiologists (AUC 0.80 and 0.88). Benign versus malignant peripheral nerve sheath tumor differentiation reached AUC 0.94, with intermediate radiomic features correlating with premalignant atypical peripheral nerve sheath tumors in neurofibromatosis type 1. Benhabib et al. demonstrated MRI radiomics distinguishing benign myxomas from malignant myxoid sarcomas in 523 cases with AUC 0.92. A combined MRI radiomic risk score plus patient age in a cohort of 176 patients achieved AUC 0.84 for benign versus malignant differentiation.
These results collectively position radiomics as a potential "digital triage" tool that could reduce unnecessary biopsies for likely benign masses and increase diagnostic confidence for masses where biopsy is technically difficult or contraindicated. The most clinically validated application remains the uterine leiomyosarcoma/leiomyoma problem, where biopsy contraindication makes a noninvasive discriminator particularly valuable.
Histologic grade assigned by the FNCLCC (Federation Nationale des Centres de Lutte Contre le Cancer) grading system is the single most important prognostic variable in soft tissue sarcoma, predicting both distant metastasis risk and disease-specific survival. However, core needle biopsy systematically underestimates the final surgical grade because it samples only a fraction of a heterogeneous tumor. High-grade regions may be missed, particularly in large tumors with central necrosis or geographic grade variation. Radiomics addresses this sampling limitation by analyzing the entire tumor volume as imaged, capturing heterogeneity that any single biopsy needle cannot.
MRI radiomics for grade prediction: Early MRI-based radiomics studies achieved high vs. low grade soft tissue sarcoma prediction with AUC 0.92, accuracy 91.4%, sensitivity 88.2%, and specificity 94.4%. Subsequent meta-analyses confirmed pooled performance with sensitivity 84%, specificity 73%, and AUC 0.91 for distinguishing low- from high-grade soft tissue sarcomas. MRI-based nomograms incorporating intratumoral habitat features and peritumoral edema outperform either radiomics alone or clinical features alone (AUC 0.868 vs. 0.856). Multicenter validation has demonstrated that some of these models can be applied across different imaging settings, an important criterion for clinical deployment. One important caveat: pooled estimates are confounded by inconsistent grade dichotomization, with some studies grouping FNCLCC grade 2 tumors with low-grade sarcomas while others classify them as high-grade.
Prognostic features beyond grade: Schmitz et al. found that heterogeneity, ill-defined margins, peritumoral edema, and peritumoral contrast enhancement were more common in high-proliferative sarcomas compared to low-proliferative sarcomas on MRI radiomics analysis. These imaging correlates of biologic aggressiveness capture tumor-host interface features invisible to standard semantic assessment.
Predicting distant metastasis and recurrence: Distant metastasis occurs in approximately 30% of soft tissue sarcoma patients and is the primary driver of sarcoma mortality. Radiomics-based machine learning methods predict distant metastasis with greater than 90% accuracy. Hu et al. built a nomogram combining MRI radiomic features and clinical data that retrospectively predicted lung metastases with AUC 0.894, substantially outperforming evaluation of tumor margins alone (AUC 0.666). IQR and kurtosis radiomic features predict prognosis and local recurrence with AUCs of 0.78-0.79, with accuracy of 86%. Multi-institutional cohort studies have also used baseline MRI radiomic features to identify postoperative progression after surgical resection.
Translating radiomics from a diagnostic and prognostic tool into one that actively guides treatment decisions is the highest-stakes application reviewed in this paper. Standard frontline systemic therapy for high-grade soft tissue sarcoma, typically doxorubicin plus ifosfamide in the neoadjuvant setting, carries significant toxicity and achieves only modest response rates in many subtypes. Identifying patients unlikely to respond before completing full neoadjuvant courses would allow earlier transition to surgery, alternative regimens, or clinical trial enrollment.
Tertiary lymphoid structures and immunotherapy: Tertiary lymphoid structures (TLS) within the tumor microenvironment are organized aggregates of immune cells that correlate with favorable responses to immune checkpoint inhibitors. Standard needle biopsy is poorly suited to detecting TLS due to sampling limitations in heterogeneous tumors. Radiomics models trained to predict the presence of intratumoral TLS from MRI could serve as a "digital biopsy," noninvasively identifying patients whose tumors have an immunologically active microenvironment. This has direct treatment implications: the PEMBROSARC trial demonstrated clinical activity of the anti-PD-1 agent pembrolizumab specifically in soft tissue sarcomas harboring TLS. If a radiomics model can reliably identify TLS-positive tumors without biopsy, it could shift clinical practice toward immunotherapy-first approaches in a subset of patients currently receiving cytotoxic chemotherapy.
Radiation treatment response assessment: For patients undergoing neoadjuvant radiation therapy before surgical resection, serial MRI provides the opportunity to assess response between treatment planning and definitive surgery. Radiomics enables quantitative comparison of pre- and post-radiation MRI data. Delta-radiomics, defined as the change in radiomic features between two time points, has been applied to identify early responders and non-responders after neoadjuvant chemotherapy. Studies have found that decreases in peritumoral edema, shape features, and texture features correlate with pathologic treatment response. ADC maps, which provide absolute measurements of tissue diffusivity directly comparable across patients and scanners without the need for normalization, represent a particularly promising substrate for response-assessment radiomics because their absolute values are physically interpretable.
MRI volumetric and texture analysis has also been applied to desmoid fibromatosis, a locally aggressive fibroblastic neoplasm treated with active surveillance, systemic agents, or radiation depending on behavior. Quantitative imaging metrics can track progression and regression of these tumors in ways that qualitative size measurements miss, particularly in tumors that show signal changes without dimensional change.
Standard oncologic response assessment in solid tumors relies on RECIST 1.1 criteria, which measure the longest diameter of target lesions on cross-sectional imaging. For soft tissue sarcomas treated with neoadjuvant chemotherapy or radiation, RECIST has well-documented limitations: many tumors that achieve extensive histologic necrosis at resection show minimal dimensional change on imaging. Conversely, tumors that increase modestly in size during therapy may still achieve pathologic complete response if the additional volume represents treatment-induced edema or hemorrhage rather than viable tumor. Radiomics offers texture- and signal-based metrics that capture these biologic response signals independent of dimensional change.
Delta-radiomics and neoadjuvant chemotherapy: Delta-radiomics computes the difference or ratio of radiomic features between pre-treatment and mid-treatment (or post-treatment) MRI acquisitions. For soft tissue sarcomas undergoing neoadjuvant doxorubicin-ifosfamide chemotherapy, changes in peritumoral edema and multiple texture features on T2-weighted MRI have been associated with pathologic treatment response. Machine learning analysis of delta-radiomic feature vectors can separate responders from non-responders earlier in the treatment course than conventional size-based assessment allows. This opens the possibility of adaptive therapy protocols in which treatment is modified for predicted non-responders before completing the full neoadjuvant course.
Case examples and visual correlation: The paper illustrates radiomics concepts with specific clinical cases. A high-grade leiomyosarcoma treated with 50 Gy neoadjuvant radiotherapy shows minimal visual change on post-treatment contrast-enhanced MRI, with persistent solid enhancement and no apparent necrosis. Histologic analysis at resection confirmed only 5% necrosis, a finding consistent with the absence of radiomic signal changes. In contrast, a pleomorphic rhabdomyosarcoma treated with combined neoadjuvant chemotherapy and radiation shows diminished peripheral enhancement thickness, captured quantitatively as a leftward shift in the pixel intensity histogram (increased skewness). A dedifferentiated liposarcoma shows subtle increase in hypointense internal components after treatment, reflecting post-treatment hyalinization and hemorrhage, changes difficult to quantify with RECIST but captured by first-order radiomic statistics. In a fourth case, a high-grade pleomorphic fibroblastic sarcoma treated with doxorubicin-dacarbazine plus radiation shows markedly decreased tumor enhancement with increased skewness, corresponding to 98% histologic tumor necrosis at resection.
ADC-based radiomics for undifferentiated pleomorphic sarcoma: Valenzuela et al. expanded ADC map analysis beyond simple standard deviation measurements to full radiomic feature extraction for undifferentiated pleomorphic sarcomas of the extremity, demonstrating the ability to predict pathologic treatment response from diffusion-weighted MRI. Because ADC values represent absolute physical tissue properties rather than scanner-relative signal intensities, ADC-based radiomic features may generalize better across imaging sites than conventional T1 or T2 features.
Validation gaps: A recent systematic review found that only 59% of soft tissue sarcoma radiomics studies assessed feature reproducibility, only 69% conducted internal validation, and fewer than 25% included external validation from an independent institution or imaging site. This creates a body of literature in which the majority of published models have not been tested on any patient population other than the one used for training. The consequence is systematic overestimation of real-world performance. Models that achieve AUC 0.92 in internal validation commonly show performance degradation of 5-15% when externally validated on multicenter cohorts.
Technical variability across scanners: Soft tissue sarcoma patients undergo serial imaging for staging, response assessment, and surveillance, often across different institutions with different scanner models, field strengths, acquisition protocols, slice thicknesses, and contrast phases. These technical variations confound radiomic feature extraction because many texture and first-order features are sensitive to voxel size, signal-to-noise ratio, and reconstruction kernel. Two technical solutions are required: image resampling, which harmonizes spatial resolution and voxel size across datasets, and data resampling, which addresses the class imbalance problem arising from the rarity of specific sarcoma subtypes, using oversampling (e.g., SMOTE) or undersampling approaches to improve machine learning model performance on minority classes.
Multi-sequence integration complexity: Modern sarcoma MRI protocols include T1, T2, fat-suppressed T2 (STIR), post-contrast T1, DWI, and sometimes DCE sequences. Each provides complementary biological information, but integrating features across sequences introduces methodological choices without consensus standards. Studies use three competing approaches: feature-level fusion (concatenating all features into a single pool), model ensemble methods (training separate models per sequence then combining outputs via weighted summation), and attention-based channel weighting (learning a single cross-sequence representation). Without standardization of these approaches, results across studies remain difficult to compare and replicate.
Feature dimensionality and small datasets: Many sarcoma subtypes are individually rare even at specialized referral centers, limiting training dataset size to dozens to a few hundred cases. Radiomics pipelines typically extract hundreds to thousands of features from multi-sequence MRI, routinely exceeding sample size. This "curse of dimensionality" makes overfitting likely unless strict feature reduction is applied via ICC filtering, correlation-based redundancy removal, LASSO, mRMR, or PCA. Additionally, the lack of reimbursement and relative value unit (RVU) credit for radiomics analysis in clinical practice creates a structural disincentive for radiologists to incorporate it into routine workflow outside of research settings.
Automated segmentation with foundation models: The bottleneck of manual tumor segmentation, which introduces interobserver variability and limits throughput, is being addressed by AI-based segmentation tools. Foundation models such as the Segment Anything Model (SAM) and its medical derivative MedSAM offer region-of-interest delineation across imaging modalities and institutions with reduced manual input. These models, pre-trained on large and diverse image datasets, can be fine-tuned for sarcoma-specific segmentation tasks and promise to reduce the labor and variability that have limited the scalability of radiomics research and deployment.
Hybrid deep learning-radiomics: An emerging approach combines hand-crafted radiomic features with high-level latent representations learned by deep convolutional or transformer neural networks. In these hybrid architectures, the interpretability of classical radiomic features is retained alongside the representational power of deep learning, potentially yielding models that are both more accurate and more explainable than either approach alone. Deep learning networks trained end-to-end on tumor image patches can capture spatial patterns that fall outside the feature vocabulary of conventional radiomics, including architectural patterns at the tumor-stroma interface that correlate with immunologic microenvironment composition.
Real-time PACS integration: The authors envision a future reading room workflow in which a patient's current imaging study is automatically segmented in real time, radiomic features are extracted, and comparison with prior imaging produces a quantitative report of tumor response or progression that the radiologist incorporates into their dictation. For a patient presenting with a new soft tissue mass, PACS-integrated AI could provide differential diagnoses and malignancy probability scores alongside the standard radiologic report. This workflow would require PACS vendor support, regulatory clearance, and prospective validation that the AI-augmented reports improve clinically meaningful outcomes.
Multi-omics integration and health economics: Future studies should explore associations between imaging-derived radiomic biomarkers and other "-omics" data layers including genomics, transcriptomics, and metabolomics. Integrating imaging phenomics with molecular data could reveal mechanistic links between radiomic signatures and tumor biology, providing interpretable rationales for radiomic predictions. Separately, health economic evaluation will be essential before widespread clinical adoption: cost-effectiveness studies need to quantify the downstream clinical value of radiomic-guided decisions, including avoided biopsies, more appropriate treatment selection, and earlier detection of progression, to justify the infrastructure and workflow investment required for PACS integration.