Sarcomas are a rare and biologically heterogeneous group of malignant tumors arising from mesenchymal tissue, which includes muscle, fat, bone, cartilage, blood vessels, and fibrous connective tissue. In the United States they account for just 1% of new cancer cases and cancer-related deaths, but that aggregate statistic obscures a devastating burden on younger patients: among children and adolescents between ages 0 and 19, sarcomas are responsible for 13% of all cancer-related deaths. The incidence is also somewhat higher in males, particularly for childhood sarcomas, although the biological basis of this sex difference remains poorly understood.
The diagnostic delay problem: Clinical signs of sarcoma are frequently nonspecific, mimicking benign musculoskeletal conditions, which creates dangerous diagnostic delays. Adding to the challenge is the sheer rarity of the disease: general practitioners encounter, on average, only one soft tissue sarcoma case over 24 years of practice. As a result, these tumors frequently reach a specialist only after months of mismanagement. The stakes of a late diagnosis are quantified clearly in the literature: patients with soft tissue sarcomas larger than 5 cm at diagnosis carry a relative risk of death 2.1 times higher than those with tumors under 5 cm (Pisters et al.).
Current imaging and its limits: Radiological imaging is central to the diagnostic workup. For bone sarcomas, radiography, MRI, and CT are the primary tools; for soft tissue lesions, ultrasound and MRI dominate. Yet image interpretation remains inherently subjective, dependent on the individual radiologist's experience. Histopathology remains the gold standard for definitive diagnosis, but it requires an invasive biopsy procedure and is subject to its own interobserver variability. Genetic profiling has emerged as a valuable complement, identifying subtype-specific mutations and molecular signatures, but it is expensive and not universally available.
This paper presents a machine learning framework built on radiomic features extracted from MRI scans, designed to classify sarcoma versus healthy tissue without requiring invasive procedures. The approach deliberately prioritizes explainability alongside accuracy so that clinicians can interrogate the model's reasoning and integrate it as a second opinion in existing diagnostic workflows, rather than treating it as an opaque black box.
The study draws on the IceSG (Icelandic Sarcoma Group) dataset, a collection of sarcoma cases assembled by Landspítali University Hospital spanning a 20-year period from 2004 to 2023. The cohort comprises 86 patients diagnosed with various bone and soft tissue tumors, with ethics approval from both the Landspítali Ethics Committee (no. 5/2024) and the Scientific Research Committee (no. 93/2024). The average age at diagnosis was 48.78 years (standard deviation 23.52), reflecting a wide age distribution from 9 to 108 years, which is clinically representative of the full range of sarcoma presentations across the lifespan.
Sarcoma subtype distribution: Soft tissue sarcomas were the most common diagnosis, accounting for 54 patients. Osteosarcoma was the second most frequent at 14 cases, followed by chondrosarcoma and Ewing sarcoma with 7 cases each. Less common diagnoses included dermatofibrosarcoma protuberans (1 case) and giant cell tumor of bone (2 cases). Because each patient underwent imaging in multiple anatomical planes, the dataset yielded 186 total MRI exams, providing a larger effective sample than the patient count alone suggests.
MRI sequence heterogeneity: A defining characteristic of this dataset, and one of its most important strengths, is the heterogeneity of the imaging sequences included. Both T1-weighted and T2-weighted sequences were incorporated, with and without contrast enhancement, with and without fat suppression, and including STIR (Short Tau Inversion Recovery) sequences. Scans were acquired in axial, coronal, and sagittal planes. The total breaks down as 95 T1 exams and 91 T2 exams. This deliberate inclusion of real-world variability in acquisition parameters, scanner generations, and anatomical locations is unusual in the radiomics literature, where most studies use controlled single-protocol datasets. The authors specifically included the single STIR exam because its visual appearance closely resembles T2 with fat saturation, consistent with methodological precedent from Casale et al.
Healthy tissue sampling strategy: Since no healthy control subjects were available, the researchers extracted healthy tissue regions from the contralateral limb of the same sarcoma patients. This intra-patient strategy ensures that healthy and tumor samples share the same imaging context, scanner, and acquisition parameters, effectively controlling for site-level and hardware-level confounders. The approach yielded 175 healthy-tissue samples, producing a dataset of 361 total samples that is approximately balanced between the two classes (186 sarcoma, 175 healthy).
The analytical pipeline began with manual segmentation of tumor and healthy tissue regions using Mimics v25.0 software (Materialise, Leuven, Belgium). Each segmented region was labeled "Sarcoma" or "Healthy," with up to three image sets per patient corresponding to axial, coronal, and sagittal planes. From these segmented volumes, radiomic features were extracted following the standardized PyRadiomics pipeline, which ensures reproducibility and comparability with other studies using the same framework.
Why handcrafted features over deep features: The authors made an explicit decision to use handcrafted (engineered) radiomic features rather than deep learning features. The rationale is methodologically sound: deep learning architectures typically require large-scale training datasets to learn meaningful representations, while handcrafted features can characterize tissue properties robustly even with limited data. Given that the entire dataset comprised 361 samples across a rare cancer type, this trade-off was appropriate.
Wavelet transforms for higher-order features: To enrich the feature set, three-dimensional wavelet transforms were applied to each MRI volume. The 3D wavelet decomposition generates eight distinct frequency sub-bands by applying combinations of low-pass (L) and high-pass (H) filters along the X, Y, and Z axes: LLL (global structure), LLH, LHL, LHH, HLL, HLH, HHL, and HHH (fine spatial detail, edges, and textures). From each sub-band, 93 intensity-based and texture features were extracted. Shape-based features were excluded from wavelet sub-bands because they depend only on the spatial mask geometry and not on voxel intensity values.
Feature count breakdown: The original image contributed 107 features (93 intensity and texture features plus 14 shape-based features). The eight wavelet sub-bands each contributed 93 features, adding 744 features (93 x 8). The combined feature set totaled 851 radiomic features per segmented region. This high-dimensional feature space captures first-order statistical descriptors (mean, median, skewness, energy), texture features from Gray Level Dependence Matrix (GLDM), Gray Level Run Length Matrix (GLRLM), and Gray Level Co-occurrence Matrix (GLCM), as well as shape descriptors encoding tumor geometry and volume.
The classification model chosen was Random Forest (RF), an ensemble method that constructs multiple decision trees during training and combines their outputs for a final prediction. Random Forest was selected for its ability to handle high-dimensional feature spaces, its natural resistance to overfitting through bagging (bootstrap aggregating), and its built-in feature importance mechanism. The classifier was implemented using scikit-learn v1.5.1 in Python 3.9, with supporting libraries pandas v2.2.1 (data manipulation) and numpy v1.26.4 (numerical operations).
Dataset partitioning: Before any model training, a stratified hold-out test set of 58 samples (28 healthy, 30 sarcoma) was reserved and sealed until final evaluation. This test set was never used during hyperparameter tuning or model selection, ensuring an unbiased performance estimate. The remaining samples formed the development set used for all optimization steps. Stratified partitioning preserved class proportions across both subsets.
Nested cross-validation framework: To avoid information leakage between model selection and model evaluation, the authors implemented a Nested Cross-Validation (NCV) strategy. In NCV, an outer loop estimates generalization performance while an inner loop performs hyperparameter optimization. Both loops used stratified k-fold partitioning to maintain consistent class distributions. After testing standard k values and finding that they did not yield consistent parameter selection, the authors selected a specific k for the outer loop that produced stable results across folds (the exact k is described in the paper). This is a methodologically rigorous choice that explicitly prevents the optimistic bias that arises when hyperparameter tuning and performance estimation share the same data.
Hyperparameter grid: The inner loop performed Grid Search over three Random Forest hyperparameters: n_estimators (100, 200, 1000), max_depth (None, 5, 10), and min_samples_split (2, 5). The most frequently selected configuration across outer folds was max_depth = None (unrestricted tree growth), min_samples_split = 2 (split whenever at least 2 samples are present), and n_estimators = 200 (200 trees in the ensemble). Models were evaluated within each fold using accuracy, precision, recall, F1-score, and ROC-AUC. The final model was retrained on the entire development set using these optimal hyperparameters before evaluation on the sealed hold-out test set.
On the sealed hold-out test set of 58 samples, the final Random Forest model achieved an accuracy of 0.724, precision of 0.719, recall of 0.767, F1-score of 0.742, and ROC-AUC of 0.871. The authors specifically prioritize the F1-score as the primary metric for this medical imaging task because it is the harmonic mean of precision and recall, penalizing both false positives (healthy tissue classified as sarcoma, potentially leading to unnecessary interventions) and false negatives (sarcoma missed, with potentially fatal consequences). An F1-score of 0.742 on a heterogeneous dataset combining multiple sarcoma subtypes, imaging protocols, scanner generations, and acquisition planes represents a meaningful clinical signal.
Cross-validation performance across folds: During nested cross-validation on the development set, performance varied across outer folds in a manner that reflects genuine dataset variability rather than a single optimistic split. Fold-level F1-scores ranged from approximately 0.757 to 0.848, and ROC-AUC ranged from 0.768 to 0.938. The variability across folds underscores how challenging the classification task is when imaging conditions differ substantially, and it argues against reporting only best-case or average results without confidence intervals.
Comparative benchmarking: The authors also evaluated seven other classifiers on the same balanced binary dataset using WEKA v3.8.6. The Random Forest model outperformed all alternatives on F1-score (0.742) and ROC-AUC (0.871). The next-best competitors were LWL (Locally Weighted Learning, accuracy 0.741, ROC-AUC 0.740) and SVM (accuracy 0.724, ROC-AUC 0.718). RandomTree achieved an F1-score of 0.653 and ROC-AUC of 0.700. The IBK (k-Nearest Neighbors) classifier obtained a relatively high recall of 0.741 but low precision of 0.571, indicating over-prediction of sarcoma. KStar failed entirely, assigning every sample to the negative class and producing no true positives. Naive Bayes and BayesNet both showed a tendency to over-predict the positive class, yielding imbalanced precision-recall trade-offs.
The ROC-AUC of 0.871 for the Random Forest is particularly notable given the heterogeneous dataset. AUC reflects the model's ability to rank sarcoma cases above healthy cases across all possible classification thresholds, and 0.871 indicates strong discriminative capability even when trained on mixed protocols, planes, and subtypes from a single center.
A central contribution of this paper is its dual-layer approach to explainability, combining global feature importance analysis with local, instance-level explanations via LIME (Local Interpretable Model-agnostic Explanations). The authors justify this design by emphasizing that clinical adoption of AI in radiology requires not just high accuracy but transparent reasoning that aligns with known tissue biology. A model clinicians cannot interrogate is a model they will not trust, particularly for rare cancers where expert intuition is hard-won.
Global interpretability via feature importance: Random Forest provides a built-in global interpretability mechanism through the feature_importances_ attribute, which quantifies each feature's average contribution to reducing impurity across all decision trees. Of the 851 features in the model, 849 contributed at least marginally to classification decisions. The top 10 most influential features were heavily concentrated in wavelet-derived descriptors, particularly from the LHL and HLL decomposition levels. The single highest-ranked feature was wavelet-LHL_firstorder_Mean (importance score 0.0249), followed by wavelet-HLL_firstorder_Median (0.0158) and wavelet-HLL_firstorder_Mean (0.0157). Among shape/texture features, original_gldm_LargeDependenceLowGrayLevelEmphasis (0.0132) and original_gldm_LowGrayLevelEmphasis (0.0102) were highly ranked, indicating that low-intensity voxel regions and their spatial dependencies carry strong discriminative power. GLRLM features (low gray-level run emphasis, short run low gray-level emphasis) also appeared in the top 10.
Biological interpretation: The dominance of wavelet-derived first-order statistics (mean, median, skewness) and low gray-level emphasis features aligns with known MRI characteristics of sarcoma: these tumors often present with heterogeneous signal intensity, irregular texture, and regions of low voxel intensity corresponding to necrosis, calcification, or fibrous stroma. The fact that texture patterns from wavelet sub-bands capturing specific spatial frequency combinations (LHL: low-pass X and Z, high-pass Y; HLL: high-pass X, low-pass Y and Z) are most informative suggests that directional intensity gradients in particular anatomical axes are diagnostically relevant.
Local interpretability via LIME: LIME explains individual predictions by approximating the Random Forest's local behavior with a simpler, interpretable linear model. The authors applied LIME to four representative test cases covering all prediction outcomes: true negative (healthy correctly classified), true positive (sarcoma correctly classified), false positive (healthy misclassified as sarcoma), and false negative (sarcoma misclassified as healthy). In the true negative case, all top 10 features had negative contributions, consistently directing the prediction toward the healthy class. In the true positive case, 9 of 10 top features had positive contributions, indicating a confident and coherent sarcoma prediction. For both misclassification cases, the LIME analysis revealed mixed feature contributions with no single dominant direction, highlighting how overlapping feature distributions between classes create ambiguous decision zones. Individual feature weights were small (in the range of a few hundredths), expected given that 849 features collectively drive the prediction rather than any single variable dominating.
Small dataset with single-center origin: The most significant limitation is the dataset size of 86 patients and 361 samples from a single tertiary referral center in Iceland. While the authors expand the effective sample size by using multiple imaging planes per patient (yielding 186 sarcoma exams), and while the dataset's heterogeneity across 20 years of acquisition is a genuine strength for robustness, 86 patients across multiple sarcoma subtypes means that some subtypes (chondrosarcoma, Ewing sarcoma, 7 cases each; dermatofibrosarcoma protuberans, 1 case) are represented by very few examples. The authors are direct about this constraint: generalizability to broader patient populations, different referral patterns, and different national imaging standards cannot be assumed without external validation.
Manual segmentation as a bottleneck: All tumor and healthy-tissue volumes were manually segmented using Mimics v25.0. Manual segmentation is the current standard for radiomic feature extraction, but it is labor-intensive, not scalable, and introduces interobserver variability. Different radiologists segmenting the same tumor may draw boundaries that differ at the margins, and because radiomic features are sensitive to the exact contour of the segmented region, these differences can propagate into feature values and affect classifier outputs. The authors acknowledge that future work should investigate automated segmentation methods, potentially using deep learning-based approaches such as nnU-Net, to remove this bottleneck and reduce inter-operator variability.
Classification scope: binary only: The current framework is binary, distinguishing sarcoma from healthy tissue. It does not attempt subtype classification between soft tissue sarcoma, osteosarcoma, chondrosarcoma, and Ewing sarcoma, and it does not address grading (low-grade versus high-grade), staging, or treatment response prediction. These are all clinically critical downstream questions. The comparatively simpler binary classification task was appropriate for this proof-of-concept study, but extending the framework to multi-class subtype discrimination or grading would require substantially larger datasets and potentially different modeling architectures.
Deep learning context: The authors acknowledge the rapid progress of deep learning for sarcoma imaging analysis. Related work on the same IceSG dataset (Correra et al.'s prior slice-based study) achieved 76.02% accuracy using a different experimental setup with leave-one-out cross-validation on individual slices, making direct comparison difficult. Studies applying CNNs for deep feature extraction (Dai et al., 172 patients, uterine sarcoma vs. atypical leiomyoma) and fully learned architectures demonstrate the active trajectory of the field, and the authors position their explainable radiomics approach as complementary rather than competing with deep learning, particularly for data-limited settings.
Multimodal data fusion: The authors identify multimodal AI integration as the most promising direction for expanding this framework. Sarcoma diagnosis and treatment planning currently draws on MRI imaging, histopathology, molecular profiling, genetic data (including fusion gene detection, copy number alterations, and subtype-defining translocations), and clinical variables such as tumor size, anatomical location, and patient age. Machine learning systems that integrate radiological features with transcriptomic and genomic data, sometimes called radiogenomics, have already demonstrated improved prognostic performance in soft tissue sarcoma in published work (Integrating Radiogenomics and Machine Learning in Musculoskeletal Oncology Care). The computational challenge is designing fusion architectures that can handle different data types, different levels of missing data across modalities, and the small cohort sizes typical of sarcoma research.
Generative models for data augmentation: A research direction with direct relevance to the dataset size problem is the use of Generative Adversarial Networks (GANs) to synthesize realistic sarcoma MRI images for training augmentation. Published work (Prediction of soft tissue sarcoma response to radiotherapy using longitudinal diffusion MRI and a deep neural network with GAN-based data augmentation) has demonstrated feasibility in sarcoma-adjacent settings. Synthetic data generated from real patient images can effectively expand training datasets for rare subtypes, though it requires careful validation to ensure the synthetic images faithfully reproduce the statistical properties of real tumors rather than introducing artifacts or biases.
Automated segmentation: Removing the dependency on manual segmentation is a prerequisite for clinical scalability. Automated deep learning segmentation models, particularly nnU-Net, have achieved performance approaching manual delineation for numerous organ and tumor types. Applying such models to sarcoma segmentation across the full range of MRI protocols used in clinical practice, including the heterogeneous T1, T2, STIR, fat-suppressed, and contrast-enhanced sequences in this dataset, remains an active research challenge but is technically achievable with sufficient annotated training data.
Clinical integration pathway: The paper frames the ultimate goal as a decision-support tool that operates alongside radiologists and clinical teams rather than replacing their judgment. In this model, the AI system produces a classification output with an associated confidence score and a LIME-style explanation identifying the specific MRI regions and feature values driving the prediction. The clinician can then weigh this information against biopsy results, patient history, and imaging gestalt. This second-opinion design lowers the barrier to adoption because it positions the tool as an assistant rather than a gatekeeper, while still providing meaningful uncertainty quantification through the ROC-AUC and confusion matrix outputs.