Integration of pre-treatment computational radiomics, deep radiomics, and transcriptomics enhances soft-tissue sarcoma patient prognosis

Soft-tissue sarcomas (STS) are heterogeneous malignant mesenchymal tumors with diverse clinical, radiologic, histologic, and molecular features that influence prognosis. Contrast-enhanced MRI is the preferred modality for locally advanced STS and reveals distinct radiologic phenotypes with substantial intra- and inter-tumor heterogeneity. Qualitative assessment of MRI features such as heterogeneous T2 signal, necrosis, and peritumoral enhancement has been associated with FNCLCC histologic grade, metastasis-free survival (MFS), and overall survival (OS). However, qualitative assessment is subjective and limited. Radiomics and deep learning offer quantitative, reproducible methods to capture tumor heterogeneity. This study investigates whether pre-treatment handcrafted and deep radiomics can identify clinically meaningful STS subgroups, their relationships to histopathology and transcriptomics, and their prognostic value for MFS, as well as whether combining radiomics with transcriptomics improves prognostication.

Prior work has shown that conventional MRI features correlate with FNCLCC grade and outcomes in STS, but qualitative assessments are subjective. Multiple studies have explored MRI-based radiomics to predict grade, treatment response, and outcomes, demonstrating potential prognostic value. Transcriptomic signatures have also been reported to improve prognostication in STS and other cancers. Radiogenomic approaches suggest complementary information between imaging-derived features and molecular profiles. Nonetheless, comprehensive integration of handcrafted radiomics, deep radiomics, and transcriptomics in locally advanced STS for prognostication has remained limited, motivating the present study.

Study design and cohort: Single-center retrospective study approved by the institutional review board, with informed consent waived. Consecutive adults with newly diagnosed, locally advanced STS treated with curative intent between May 2008 and May 2020 were identified from a surgical database. Inclusion required pre-treatment MRI with gadolinium-enhanced sequences and histopathologic confirmation by expert sarcoma pathologists. A total of 225 patients with adequate MRI were included for radiomics; a subset of 110 had suitable pre-treatment tissue for RNA sequencing.

Clinical and pathological data: Collected variables included age, sex, WHO performance status, tumor depth, location, longest diameter, histologic type, and FNCLCC grade (from biopsy or surgical specimen), treatments (radiotherapy, chemotherapy), and surgical margins (R0/R1/R2).

MRI acquisition: Exams acquired on multiple 1.5T systems. Protocols included T1-weighted (pre-contrast), T2-weighted, and contrast-enhanced fat-suppressed T1-weighted sequences. Repetition/echo time ranges: T1-WI 500–700/10–15 ms, T2-WI 2400–4500/70–130 ms. In-plane resolution 0.75×0.75 to 1.4×1.4 mm²; slice thickness 1–7 mm.

Semantic radiology: Two senior radiologists performed consensus semantic evaluation. A high-risk radiophenotype was assigned when at least two of: heterogeneous T2 signal, necrosis (fluid-like T2 hyperintensity without enhancement), and peritumoral enhancement were present; otherwise low-risk.

Handcrafted radiomics (h-RF): MRI post-processing used ITK: resampling to 1×1×4 mm, bias correction, and histogram matching. Tumors were manually segmented on CE-T1-W and propagated to T1-W and T2-W with adjustments. Features extracted with LIFEx (IBSI-compliant): first-order, second-order textures (GLCM, GLRLM, NGLDM, GLZLM), and shape features, totaling 175 h-RFs. Reproducibility assessed on 30 randomly selected cases. Unsupervised consensus clustering of patients based on h-RFs used iterative K-means with bootstrap/hierarchical clustering (Pearson distance, average linkage) and resampling.

Deep radiomics (d-RF): Tumor-centered CE-T1-W slices were preprocessed; segmentation masks defined regions. Two autoencoder architectures were trained: convolutional autoencoder (CAE) and half-supervised CAE (HSCAE). Models were trained on 200 patients (Training cohort) with repeated cross-validation; 25 were held out (Testing cohort). Reconstruction performance assessed via mean square error (MSE), which remained <1% in both training and testing sets. Latent features (deep radiomics) were extracted (reported dimensionality 104 per image) and subjected to unsupervised hierarchical clustering (Pearson distance, average linkage) to define CAE and HSCAE groupings. For the 25-patient test set, cluster assignment used a centroid-distance method.

Transcriptomics: For 110 patients, pre-treatment frozen or FFPE tissue underwent RNA sequencing. Reads were quality-controlled and aligned; gene-level counts derived from uniquely aligned reads. Batch effects were corrected and expression normalized (as described in Supplementary Methods). Unsupervised consensus clustering on RNA-seq data identified transcriptomic groups.

Statistical analysis: Associations among radiomics groups (h-RF, CAE, HSCAE), transcriptomic groups, histologic types/grade, SARCULATOR risk groups, and semantic radiophenotype were tested using chi-square tests. MFS was analyzed with Kaplan–Meier and log-rank tests. Univariable and multivariable Cox regressions estimated hazard ratios (HRs) with 95% CIs, adjusting for SARCULATOR covariates and management factors (age, size, histology category, chemotherapy, radiotherapy, surgical margins). To evaluate complementarity, Harrell’s c-index was computed in 5-fold cross-validation for models including h-RF, HSCAE, and transcriptomics groups alone and in combination (with/without interaction), with bootstrap comparison of differences. A stepwise backward Cox approach (AIC minimization) identified the most prognostic deep clusters.

Gene-expression analyses: Differential gene expression (DGE) was performed between the combined radiomics-transcriptomics risk groups (worst vs others), using t-tests per gene, fold-change threshold of 2, and Benjamini–Hochberg adjustment. Gene set enrichment used the Molecular Signatures Database (MSigDB). A PAM-based classifier (nearest-centroid approach) was explored for discriminant signatures.

• Cohort: 225 patients (53.3% men), median age 62 years; 110 had RNA-seq.
• Unsupervised clustering from h-RFs identified three groups: A-RF (71/225; 31.6%), B-RF (87/225; 38.7%), and C-RF (67/225; 29.8%). Deep radiomics (CAE and HSCAE) consistently yielded two groups each (A vs B) in training and validation.
• Reconstruction errors for CAE/HSCAE were low (MSE <1%) in both training and testing cohorts.
• Radiomics clusters were significantly associated with clinicopathologic and radiologic features:
  • Associations with FNCLCC grade (P-values approximately 0.016–0.027), with higher proportions of grade III in A-CAE and A-HSCAE vs their B counterparts (e.g., 58.7–58.8% vs 41.7–41.9%).
  • h-RF clustering associated with histological grade (P = 0.011) and histologic types; higher rates of undifferentiated pleomorphic sarcoma in higher-risk imaging groups.
  • Radiomics clusters associated with high-risk semantic radiophenotype and SARCULATOR groups (all strong associations reported; several P < 0.001).
• Radiomics clusters were not significantly associated with transcriptomic clusters, indicating complementary information.
• Metastatic relapses occurred in 102/225 (45.3%) patients; 5-year MFS probabilities were approximately 78.9% and 66.4% at specified time points (as reported).
• Survival analyses (selected results):
  • h-RF groups: Compared to the reference group, the B-RF group showed worse MFS; multivariable HR 2.84 (95% CI 1.23–6.57), P = 0.0146.
  • HSCAE groups: A-HSCAE associated with worse MFS vs B-HSCAE; multivariable HR 2.73 (95% CI 1.37–5.12), P = 0.0043; log-rank P < 0.0001.
  • In the subset with both modalities (n = 110), transcriptomic group was prognostic (univariable HR for A vs B: 2.60, 95% CI 1.12–6.02; P = 0.0261). HSCAE A-group had HR 3.90 (95% CI 1.60–9.52; P = 0.0028) vs reference.
  • Multivariable Cox including h-RF and transcriptomics: both remained associated with lower MFS (e.g., transcriptomics A HR 3.50, 95% CI 1.32–9.17, P = 0.0117; transcriptomics B HR 4.23, 95% CI 1.54–11.65, P = 0.0052). Including HSCAE and transcriptomics: transcriptomics A HR 2.93 (95% CI 1.25–6.83; P = 0.0129) and HSCAE A HR 4.28 (95% CI 1.74–10.50; P = 0.0015).
• Prognostic performance (5-fold CV c-index):
  • Transcriptomics alone: c-index 0.603 (95% CI 0.574–0.675).
  • h-RF alone: c-index 0.633 (95% CI 0.599–0.771).
  • Transcriptomics × h-RF: c-index 0.646 (95% CI 0.643–0.820); significantly higher than transcriptomics alone (P = 0.0308) and h-RF alone (P = 0.0469).
  • HSCAE alone: c-index 0.661 (95% CI 0.585–0.702).
  • Transcriptomics × HSCAE: c-index 0.709 (95% CI 0.651–0.788); higher than HSCAE alone and radiomics alone.
• Gene-expression: The poor-prognosis combined group (A transcriptomics × A HSCAE) showed significant differential expression (over 1,000 DEGs reported) and enrichment of pathways related to inflammatory response, hypoxia, epithelial–mesenchymal transition, apoptosis inhibition, cell-cycle regulation (G2M checkpoint, E2F targets), UV response, and xenobiotic metabolism.

The study demonstrates that pre-treatment MRI-based radiomics, both handcrafted and deep, stratify STS into subgroups that correlate with known adverse clinical, histological, and radiologic features and predict metastatic relapse-free survival. Although radiomics clusters do not align with transcriptomic clusters, each modality independently captures prognostic aspects of tumor biology. This lack of direct correspondence suggests complementarity rather than redundancy between macro-scale imaging features and micro-scale gene-expression profiles. Combining radiomics (particularly HSCAE-derived deep features) with transcriptomics significantly improves prognostic discrimination over either alone, as evidenced by higher c-indices. The enriched molecular pathways in the worst combined group (inflammation, hypoxia, EMT, cell-cycle activation) are consistent with aggressive tumor behavior and may explain the radiologic phenotypes captured by deep radiomics. These findings support a multi-omics approach to risk stratification in STS and indicate that deep radiomics can enhance interpretation and prognostic power of transcriptomic signatures.

Pre-treatment handcrafted and deep radiomics from MRI identify clinically meaningful STS subgroups associated with histology, grade, and radiologic risk features and predict MFS. Importantly, radiomics and transcriptomics provide complementary prognostic information; their integration yields significantly improved prognostic performance compared with either modality alone. The study highlights the potential for a radiomic–transcriptomic prognostic signature to refine risk stratification in STS and inform patient management. Future work should involve prospective, multi-center validation on larger cohorts, standardization of MRI acquisition and segmentation, exploration of additional imaging sequences or modalities, and development of clinically deployable, interpretable multi-omics models.

• Single-center, retrospective design may limit generalizability and introduce selection biases.
• Small independent testing cohort (n = 25) limits robustness of deep-learning validation; external validation is needed.
• Heterogeneity of MRI scanners and acquisition parameters despite post-processing may introduce variability.
• Manual segmentation by a single radiologist may introduce observer bias, though reproducibility was partially assessed on a subset.
• Incomplete molecular data (only 110/225 with RNA-seq) may affect combined analyses.
• Code not publicly available; reproducibility depends on access upon request.
• Immediate clinical applicability is limited; further standardization and validation are required.