Chondrosarcomas (CS) account for 20 to 30 percent of all primary bone tumours in adults. Most arise from the medullary cavity and are classified as primary central conventional chondrosarcomas, which are then graded by histopathology from low to high. In long bones, the 2020 World Health Organization classification introduced a critical distinction: low-grade (grade I) cartilaginous lesions in the appendicular skeleton are now designated atypical cartilaginous tumours (ACT), not chondrosarcoma, because they are locally aggressive but have an indolent clinical course and extremely low metastatic potential. High-grade CS (grades II and III) of long bones, by contrast, carry substantial recurrence rates after surgery and meaningful metastatic risk.
Treatment divergence: The therapeutic consequences of this distinction are profound. For high-grade CS, wide surgical resection with negative margins remains the standard of care, since there is no effective chemotherapy or radiation for this histotype. ACT, however, is now increasingly managed with active surveillance (watchful waiting), especially at specialized bone tumour referral centres, rather than curettage or resection. Several long-term follow-up studies have confirmed low rates of tumour growth and no documented transformation into high-grade CS at follow-up, which has progressively erased the historical justification for routine surgery in ACT patients. The shift toward surveillance avoids operative morbidity, but it requires reliable imaging follow-up, typically with MRI at intervals, which generates substantial cost and patient burden.
The diagnostic gap: Pre-operative biopsy is unreliable for cartilaginous bone tumours. Sampling errors are common because of the spatial heterogeneity of cartilaginous matrix, and interobserver discordance in tumour grading is well documented even among specialized bone tumour pathologists. Standard imaging can differentiate high-grade from low-grade lesions through features such as cortical breakthrough, periosteal reaction, and soft-tissue extension, but interobserver variability is also a documented problem among expert musculoskeletal radiologists. A reproducible, objective imaging-based grading tool would directly address this gap. Radiomics applied to the most accessible imaging modality, standard X-rays, offers the prospect of extracting quantitative features that go beyond what the human eye perceives.
This study, published in EBioMedicine in 2024, tests whether a radiomics-based Support Vector Machine (SVM) classifier trained on plain frontal-view X-rays can differentiate ACT from high-grade CS of long bones as accurately as specialist musculoskeletal radiologists who have access to MRI and CT in addition to X-rays.
This was a retrospective, IRB-approved study conducted at two tertiary bone sarcoma referral centres in Italy: IRCCS Orthopaedic Institute Galeazzi in Milan (centre 1) and IRCCS Regina Elena National Cancer Institute in Rome (centre 2). All 150 patients had surgically treated lesions with post-surgical histopathology confirming the diagnosis, making this a high-quality ground-truth dataset rather than one relying on radiological or clinical diagnosis alone. Reporting followed both the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) and the Standards for Reporting of Diagnostic Accuracy (STARD).
Inclusion criteria and exclusion criteria: To be included, a patient needed: (i) ACT or primary central high-grade CS (grade II or III) of long bones treated surgically and confirmed by post-surgical pathology; (ii) pre-operative plain radiographs performed within three months before surgery; and (iii) MRI and/or CT available to help identify tumour margins before X-ray segmentation. Exclusion criteria removed metacarpal, metatarsal, and phalangeal locations; recurrent tumours; pathological fractures; and cases of multiple enchondromatosis (Ollier disease or Maffucci syndrome), all of which would introduce heterogeneous biology or imaging artifacts.
Cohort structure: Centre 1 contributed 120 patients, split temporally into a training cohort (surgery 2011-2020, n = 95: 71 ACT and 24 high-grade CS) and an internal test cohort (surgery 2021-2022, n = 25: 19 ACT and 6 high-grade CS). This temporal split is important because it tests whether the model generalises over time under real-world drift in scanner parameters or clinical practice. Centre 2 contributed 30 patients (12 ACT and 18 high-grade CS) as a geographically independent external test cohort. Notably, the external cohort was enriched for high-grade CS (60% vs. 25% in the training cohort), which is explained by a selection bias from the unavailability of X-rays in one-third of the initially identified centre 2 patients. Imaging was acquired on different digital radiography systems at each centre: Siemens Ysio at centre 1 and Clisis Evolution (General Medical Merate) at centre 2, with substantially different image resolutions (median 2135 x 2508 pixels at centre 1 vs. 3520 x 3520 pixels at centre 2).
Tumour locations spanned femur, humerus, fibula, tibia, and radius in both centres. There was no significant difference in age (p = 0.263, Mann-Whitney U) or sex (p = 0.410, Chi-square) between the two centres, supporting the comparability of the populations for independent testing.
The radiomics pipeline followed International Biomarker Standardization Initiative (IBSI) guidelines and was executed on the Trace4Research radiomic/AI platform (DeepTrace Technologies). The workflow proceeded in four stages: segmentation, pre-processing, feature extraction, and dimensionality reduction, each with explicit methodology to ensure reproducibility.
Segmentation: A last-year radiology resident (K.N.) manually delineated an oval-shaped region of interest (ROI) encompassing each lesion on pre-operative frontal-view X-rays. Before drawing ROIs on X-rays, the reader first reviewed MRI and/or CT (whichever was available) in all patients to identify lesion margins, then transferred that spatial knowledge to the X-ray segmentation. The reader was aware the study involved cartilaginous bone tumours but was blinded to post-surgical pathology and grading, limiting confirmation bias. This approach of cross-modality-guided X-ray segmentation is pragmatically important: it mimics a real clinical workflow where advanced imaging would typically be used for initial lesion identification before a surveillance protocol relying on X-rays is established.
Pre-processing: Images were resampled to isotropic pixel spacing using a down-sampling scheme with 1 mm slice thickness and intensity was discretized using 64 fixed bins. Standardized pre-processing ensures that features computed from images of different resolutions and acquisition parameters are comparable, a critical step given the different scanners used at the two centres.
Feature extraction: Radiomic features were extracted from the segmented ROI across five feature families: morphology, intensity-based statistics, intensity histogram, and four texture families including Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Run Length Matrix (GLRLM), Gray-Level Size Zone Matrix (GLSZM), Neighbourhood Gray Tone Difference Matrix (NGTDM), and Neighbouring Gray Level Dependence Matrix (NGLDM). A total of 121 radiomic features were extracted per lesion.
Dimensionality reduction: Three sequential filters reduced 121 features to a final set of 5. First, stability analysis tested feature reproducibility under data augmentation (random segmentation manipulation and image rotation), retaining features with Intraclass Correlation Coefficient (ICC) of 0.8 or higher. Of 121 features, 89% (n = 108) passed this stability threshold. Second, features with coefficient of variation (CoV) below 0.1 were removed as insufficiently variable across the dataset (n = 17 removed). Third, mutual information analysis estimated the information shared between each remaining feature and the class label (ACT vs. high-grade CS), removing features with mutual information below 0.28 (n = 86 removed). The CoV and mutual information thresholds were determined semi-automatically by recursively lowering them from standard starting values until a minimum of five features was retained.
The machine learning classifier consisted of an ensemble of 3 sets of 100 Support Vector Machines (SVMs), each combined with principal components analysis (PCA) and Fisher discriminant ratio selection, with a majority-vote rule for final class assignment. SVMs are well-suited to small, high-dimensional datasets because they optimize the hyperplane margin between classes in a transformed feature space (the kernel space), making them less prone to overfitting than deeper models when labeled examples are limited. The ensemble architecture improves stability by averaging over multiple classifiers, each trained on slightly different feature subsets.
Class balancing with ADASYN: The training cohort was substantially imbalanced, with 71 ACT cases and only 24 high-grade CS cases, a ratio of roughly 3:1. To prevent the classifier from trivially over-predicting the majority ACT class, the Adaptive Synthetic Sampling (ADASYN) method was applied to the training data during cross-validation. ADASYN generates synthetic instances of the minority class (high-grade CS) by interpolating between existing minority-class examples in feature space, specifically creating more synthetic instances in regions where the existing minority examples are harder to classify. This increased the effective training count of high-grade CS from 24 to 71, matching the ACT count.
Tuning with nested 10-fold cross-validation: The classifier was tuned using nested 10-fold cross-validation on the training cohort only. In nested cross-validation, an outer loop estimates model performance while an inner loop tunes hyperparameters, preventing optimistic bias from using the same data for both operations. Feature normalization (range 0 to 1) was computed from training fold values and applied to both validation and test folds within each outer fold, ensuring no data leakage. The trained model was then applied without modification to the independent internal and external test cohorts.
Comparison with radiologist performance: Two musculoskeletal radiologists with 4 and 14 years of experience at tertiary bone sarcoma centres independently reviewed the imaging studies from the internal and external test cohorts respectively, blinded to pathology and radiomics results. Crucially, both radiologists had access to all available imaging (X-rays in all cases plus CT and/or MRI where available) when making their assessments, meaning the classifier was competing against humans who had more imaging information than the classifier itself. Radiologist assessments considered features such as intralesional osteolytic areas without matrix mineralization, cortical thinning or remodelling, cortical breakthrough, periosteal reaction, oedema in adjacent bone or soft tissues, and soft-tissue extension. Classifier vs. radiologist performance was compared using McNemar's test, which is the appropriate non-parametric test for paired binary classification comparisons on the same patients.
Of the 121 radiomic features extracted, 108 passed the stability threshold (ICC ≥ 0.8), confirming that the vast majority of features were reproducibly measurable on X-ray despite segmentation variation. After removing low-CoV (n = 17) and low-mutual-information (n = 86) features, exactly 5 features survived: 3 morphology features and 2 texture features from the Neighbourhood Grey Tone Difference Matrix (NGTDM) family. All 5 were statistically significant in differentiating ACT from high-grade CS (uncorrected and Bonferroni-Holm corrected p-value less than 0.005 for each, Mann-Whitney U test).
Morphology features: The three surviving morphology features were Area, Perimeter, and Perimeter-to-Area Ratio. High-grade CS lesions showed substantially larger area (median 3021.56 pixels, 95% CI 1221.87 to 4821.25) compared to ACT (median 960.58 pixels, 95% CI 795.56 to 1125.60), reflecting the tendency of more aggressive tumours to expand to larger sizes before diagnosis. High-grade CS also showed larger perimeter values (median 221.13, 95% CI 162.32 to 279.94 vs. 120.60 for ACT), but a lower perimeter-to-area ratio (0.0725 for CS vs. 0.13 for ACT), indicating that high-grade CS lesions are geometrically more compact relative to their perimeter, reflecting irregular, lobulated expansive morphology rather than the relatively smaller, rounder ACT lesions.
Texture features (NGTDM): The two surviving NGTDM features were Busyness and Strength. The NGTDM captures the relationship between each pixel and its spatial neighbourhood in terms of grey-level intensity. Busyness measures how rapidly intensity values change between adjacent pixels, reflecting local spatial heterogeneity. High-grade CS showed markedly higher Busyness (median 3.17, 95% CI 2.09 to 4.25) compared to ACT (median 0.98, 95% CI 0.75 to 1.21), consistent with the more heterogeneous internal architecture of high-grade tumours including areas of matrix necrosis, mucoid degeneration, and irregular mineralisation. Strength, conversely, was lower in high-grade CS (median 0.25, 95% CI 0.11 to 0.39) than ACT (median 0.93, 95% CI 0.74 to 1.13), reflecting differences in the magnitude and consistency of neighbourhood intensity contrasts. These textural differences align with the known histopathological features distinguishing the two tumour types on imaging.
The convergence on size and texture heterogeneity as the discriminating radiomic signature is biologically interpretable: high-grade CS tends to present as larger, more internally heterogeneous lesions than the typically smaller, more uniform ACT. Importantly, these quantitative features capture patterns that radiologists also consider qualitatively, but standardise them into reproducible measurements independent of observer experience.
The classifier was tuned on the training cohort (mean AUC = 0.75, 95% CI 0.70 to 0.80 from nested cross-validation). When applied to the two fully independent test cohorts, performance improved substantially, suggesting the tuned model was not overfit and that the selected features had robust discriminative value beyond the training data distribution.
Internal test cohort (temporally independent, n = 25): The SVM ensemble achieved 80% accuracy (20/25 correctly classified), 83% sensitivity (5 of 6 high-grade CS correctly identified), and 79% specificity (15 of 19 ACT correctly classified). The AUC on the internal test cohort was 0.93, substantially higher than the cross-validation estimate of 0.75, which likely reflects the relatively small size of the internal test cohort (25 patients, 6 high-grade CS). One high-grade CS was misclassified as ACT (false negative) and four ACT were misclassified as high-grade CS (false positives).
External test cohort (geographically independent, n = 30): Performance on the external cohort was 80% accuracy (24/30 correctly classified), 89% sensitivity (16 of 18 high-grade CS correctly identified), and 67% specificity (8 of 12 ACT correctly classified), with AUC = 0.90. The lower specificity in the external cohort reflects a higher proportion of false positives (4 ACT misclassified as high-grade CS). Two high-grade CS were missed as false negatives. The maintained sensitivity of 89% in the geographically independent cohort using completely different scanner hardware is a clinically important finding: for the screening purpose of not missing aggressive tumours, high sensitivity is the priority metric.
Comparison to radiologists: The musculoskeletal radiologist reviewing the internal test cohort (4 years experience, access to X-rays plus MRI and/or CT) achieved 88% accuracy (22/25). The radiologist reviewing the external test cohort (14 years experience, access to all modalities) achieved 80% accuracy (24/30). Neither difference vs. the machine learning classifier reached statistical significance: p = 0.683 (McNemar's) for the internal comparison and p = 0.617 for the external comparison. The classifier, operating on X-rays only, performed equivalently to specialists who had additional MRI and/or CT. This is the central finding of the study and directly supports the potential of plain-radiograph radiomics to reduce dependence on advanced imaging in this clinical context.
Retrospective design: The study is retrospective, which is a general limitation of radiomics research at this stage but is inherent to any study requiring post-surgical histopathological ground truth. A prospective design would provide higher-level evidence for clinical validity, but assembling 150 surgically treated patients with an uncommon tumour combination at two tertiary centres already represents a substantial effort. The authors note that retrospective inclusion allowed a larger patient count than a prospective study might achieve in a similar time frame.
Limited scope (cartilaginous lesions only): The model was trained exclusively on patients with confirmed cartilaginous bone tumours. It cannot be applied to X-rays showing different bone lesions or to cases without a prior diagnosis of a cartilaginous process established through other means. The method is therefore a grading tool rather than a screening or de novo detection tool. This is not a design flaw but a deliberate constraint reflecting real-world clinical workflow: MRI is already typically used for initial lesion characterization, after which a simpler surveillance modality would be valuable.
Class imbalance and selection bias: ACT was over-represented in centre 1 (75% of training cohort) and under-represented in centre 2 (40% of external cohort). The former reflects true clinical prevalence at a tertiary orthopaedic centre where ACT is far more common than high-grade CS. The latter reflects a selection bias at centre 2 from the unavailability of X-rays in one-third of initially identified patients, which disproportionately removed ACT cases. Class imbalance was addressed by ADASYN oversampling in training, but the resulting external cohort distribution (40% ACT, 60% CS) is not representative of typical practice, which partly explains the lower specificity (67%) observed there.
Manual segmentation and lateral views: Segmentation was performed manually by a single reader, introducing the possibility of interobserver variability that could affect radiomic feature values even though stability testing was performed. The study used only frontal-view X-rays; lateral-view radiographs were not included. The authors acknowledge that lateral views may provide additional discriminating information and that future work should evaluate this. Automated deep-learning segmentation is cited as a planned next step to reduce observer dependence.
Radiomics Quality Score (RQS): The study's RQS was 44%, which the authors place in context: it is in line with average values reported across radiomics studies in a systematic review, and higher than the median RQS reported in studies specifically addressing bone chondrosarcoma radiomics. Nonetheless, an RQS below 50% indicates that methodological improvements remain achievable, particularly around prospective design, test-retest reproducibility assessment with a second reader, and independent external code validation.
The clinical significance of this work lies primarily in the context of active surveillance. As major bone tumour centres move toward watchful waiting for ACT rather than routine curettage, the frequency and modality of follow-up imaging becomes a practical and economic question. MRI, while highly accurate, is expensive and resource-intensive. The authors argue that if X-rays radiomics can reliably identify the small subset of ACT cases that progress toward or mimic high-grade CS at follow-up, then X-ray examinations could safely replace or triage advanced imaging in many surveillance visits.
Comparison with previous radiomics work: This study fits within the same research group's programme of cartilaginous tumour radiomics. In an earlier study by Gitto et al., CT radiomics-based machine learning achieved 75% accuracy in differentiating ACT from high-grade CS in a geographically independent external test cohort. A subsequent study using T1-weighted MRI radiomics achieved 92% accuracy externally. The current X-rays-based result of 80% accuracy positions X-ray radiomics between CT and MRI in diagnostic performance, which is notable given that X-rays carry lower cost, lower radiation dose than CT, and far greater accessibility in both primary care and surveillance settings.
Automated segmentation: Manual X-ray segmentation guided by MRI or CT is not a scalable clinical workflow. The authors explicitly identify deep learning-based automated segmentation as the next methodological step. Automated segmentation would remove the reader-time bottleneck, standardise ROI delineation, and eliminate the practical requirement for advanced imaging to guide the segmentation itself. Multiple deep learning segmentation architectures (including U-Net variants) have already demonstrated strong performance for bone tumour delineation on CT and MRI and could be adapted for this X-ray application.
Prospective validation and integration: The authors call for future prospective investigations to verify whether these findings transfer into real clinical practice. A prospective study would need to enrol patients presenting with cartilaginous bone lesions of uncertain grade, apply the X-ray radiomics classifier at diagnosis and at each surveillance interval, and measure clinical outcomes including rate of unnecessary curettage, missed high-grade CS, and healthcare costs compared to standard MRI-based follow-up. Integration with multimodal approaches that combine X-ray radiomics with clinical variables (age, location, symptom duration) or with targeted use of MRI only when the classifier returns an equivocal result could further optimise the clinical workflow. The dataset has been made available on Zenodo to enable independent replication and future model benchmarking.