Chondrosarcoma accounts for 20 to 30% of all primary malignant bone tumors, making it the second most common primary bone malignancy after osteosarcoma. Within this family, the distinction between low-grade (Grade 1) chondrosarcoma and higher-grade (Grade 2-4) disease carries enormous treatment consequences. Low-grade lesions of the extremity bones were formally reclassified from "malignant" to "locally aggressive" and renamed atypical cartilaginous tumours (ACT). They can be managed with intralesional curettage or even watchful waiting. Higher-grade and axially located chondrosarcomas, by contrast, require wide surgical resection with free margins, an operation with substantially greater morbidity.
Survival stakes: The 10-year overall survival drops sharply with grade: 88% for ACT/G1, 62% for G2, and 26% for G3 chondrosarcoma. These figures underscore why correct preoperative grading is not merely academic. A patient incorrectly graded as low-grade may receive only curettage and face local recurrence or transformation; a patient incorrectly graded as high-grade may undergo unnecessarily radical surgery.
The biopsy problem: Preoperative biopsy has long been the reference standard, but it carries well-recognized limitations. Cartilaginous tumors are frequently large and heterogeneous, meaning a small biopsy sample may capture only the less aggressive area. Inter-observer variability in grading even among specialist bone pathologists is documented. Biopsy-tract contamination poses additional risk. These shortcomings motivate the search for imaging-based, non-invasive alternatives.
Imaging context: MRI is the method of choice for local staging, while CT and PET-CT serve for general staging and biopsy guidance. CT additionally reveals matrix mineralization patterns and cortical changes that inform diagnosis, but visual assessment alone cannot reliably differentiate ACT from higher-grade chondrosarcoma in all cases. This gap is where radiomics and machine learning enter.
This retrospective study enrolled 120 patients with histology-confirmed cartilaginous lesions across two tertiary bone tumour centres. Centre 1 (IRCCS Orthopaedic Institute Galeazzi, Milan) provided 84 patients who formed the training cohort: 55 with ACT/G1 and 29 with G2-G4 chondrosarcomas. Centre 2 (IRCCS Regina Elena National Cancer Institute, Rome) provided 36 patients constituting the independent external test cohort: 16 ACT/G1 and 20 G2-G4 chondrosarcomas. The median patient age was 52 years (IQR 45-65) in centre 1 and 57 years (IQR 46-69) in centre 2, with no statistically significant difference between cohorts (p=0.25). Sex distribution was also balanced (p>0.99).
Inclusion criteria: Patients required a confirmed histological diagnosis based on the surgical specimen, pre-biopsy CT or PET-CT performed within one month before surgery, and surgical treatment (curettage or resection). Only conventional chondrosarcomas of long bones were included; dedifferentiated G4 cases were included in the higher-grade group. Lesions with pathological fractures, secondary tumours arising from pre-existing cartilaginous lesions, recurrent disease, or beam-hardening artifacts from metal devices were excluded.
Imaging differences between centres: Centre 1 used a 64-slice CT unit (Siemens SOMATOM Emotion) with 1 mm slice thickness, while centre 2 employed a 16-slice PET-CT unit (Siemens Biograph) with 4 mm slice thickness and a 500 mm field of view. The deliberate use of CT data from different scanners and acquisition protocols in training vs. testing is notable. It means the classifier was not merely learning scanner-specific artifacts but was generalizing across institutions.
Tumour locations: In the training cohort, femur (n=40) and humerus (n=30) dominated, followed by fibula (n=9) and tibia (n=4). The external test cohort showed a similar distribution: femur (n=21), humerus (n=5), fibula (n=6), tibia (n=4). Notably, 8 ACT cases in centre 1 underwent wide resection despite their low grade, due to anatomically challenging locations or concern about biopsy sampling errors, while all ACT cases in centre 2 underwent wide resection per institutional protocol.
Segmentation was performed by a recently-boarded musculoskeletal radiologist (S.G.) using ITK-SNAP v3.6, an open-source contouring tool. The approach was bidimensional rather than volumetric: a single axial slice showing maximum lesion extension was annotated. This choice was deliberate, following recent evidence that 2D segmentation can outperform volumetric segmentation for certain tasks and is substantially faster in clinical practice. Interreader reliability was assessed in a subgroup of 30 randomly selected training patients by two additional radiology residents, meeting the minimum numerical requirements (3 readers, 30 samples) for intraclass correlation coefficient (ICC) analysis per the Koo et al. guidelines.
Image preprocessing pipeline: Before feature extraction, all images were resampled to 1x1 mm in-plane resolution to ensure correct texture feature computation. Grey level normalization was applied using z-score normalization scaled by a factor of 100, then shifted by +300 to eliminate negative pixel values that can confound texture analysis. The resulting grey level range was 0-600, excluding outliers. Bin width selection for discretization was performed exclusively on the training set using exploratory first-order parameter extraction with widths of 2, 3, 4, and 5; a width of 3 was selected as optimal (producing an average of 59 bins per scan).
Feature extraction with PyRadiomics: Using PyRadiomics v3.0.0, a total of 919 radiomic features were extracted per segmentation from each patient. These features came from the original images plus filtered versions: Laplacian of Gaussian (LoG) filters at 5 sigma levels (1, 2, 3, 4, 5 mm) and wavelet decompositions using all combinations of high/low-pass filtering on the x and y axes. Feature categories included first-order histogram statistics, 2D shape-based metrics, and texture matrices: Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), and Neighbouring Gray Tone Difference Matrix (NGTDM). This comprehensive feature set captures not just pixel intensity distributions but spatial patterns of heterogeneity that human radiologists cannot quantify visually.
Why bidimensional segmentation? Volumetric segmentation of large cartilaginous tumours is time-consuming, potentially error-prone across hundreds of slices, and its reproducibility advantage over single-slice 2D segmentation has been questioned for certain lesion types. The bidimensional approach trades some anatomical completeness for speed and interreader consistency, which is important if the method is to be adopted in busy clinical workflows.
Starting from 919 extracted features, the pipeline applied a three-stage dimensionality reduction exclusively using training cohort data, ensuring the test set had no influence on feature selection. First, ICC-based stability analysis was run using a two-way random effect, single rater, absolute agreement model. Only features with an ICC lower confidence bound of at least 0.75 were considered stable. Of the 919 features, 30% (n=275) passed this stability threshold, and none of these had low variance. Second, highly inter-correlated features were removed using a Pearson correlation coefficient threshold of 0.80, reducing the set from 275 to 26 non-collinear features. Third, of these 26, the five with the highest information gain ratio (threshold >0.35) were selected for the final model.
The five selected features: Major Axis Length (2D shape-based) from the original image; Contrast (Neighbouring Gray Tone Difference Matrix, NGTDM) from wavelet low-high pass filtered images; Short Run Low Gray Level Emphasis (Gray Level Run Length Matrix, GLRLM) from LoG-filtered images at sigma=5; Difference Entropy (Gray Level Co-occurrence Matrix, GLCM) from original images; and Inverse Difference Moment (GLCM) from LoG-filtered images at sigma=2. These features collectively capture lesion size, spatial contrast, gray-level run patterns at coarse scales, and local textural disorder.
Handling class imbalance with SMOTE: In the training cohort, ACT (n=55) outnumbered G2-G4 chondrosarcoma (n=29) by a ratio of roughly 2:1. To address this, the Synthetic Minority Oversampling Technique (SMOTE) was applied to the training set, generating synthetic chondrosarcoma instances until both classes had 55 samples. SMOTE creates new minority-class examples by interpolating between existing ones in feature space, avoiding simple duplication. Critically, SMOTE was applied only to the training set; the external test cohort was left untouched.
LogitBoost classifier: The final classifier was a boosted linear logistic regression using the LogitBoost algorithm, trained and validated on the training cohort via 10-fold cross-validation, then applied to the external test cohort. Analysis used Weka v3.8.4, R, and scikit-learn. The Brier score (a calibration metric) and calibration curves were computed for the external test set to assess not just discrimination but also the reliability of the model's probability outputs. A Radiomics Quality Score (RQS) was calculated to benchmark methodological rigor against community standards.
In the training cohort, the LogitBoost classifier achieved 81% overall accuracy (89/110 correct, after SMOTE balancing to 110 total instances), with an area under the ROC curve (AUC) of 0.89. Breaking down by class: accuracy in identifying ACT was 84% (46/55 correct), while accuracy in identifying G2-G4 chondrosarcoma was 78% (43/55 correct). The weighted average F-score was 0.809, and the area under the precision-recall curve (PRC) was 0.895. The 10-fold cross-validation process confirmed that the model was not merely memorizing training data.
External test cohort performance: On the 36-patient independent external cohort, overall accuracy dropped to 75% (27/36 correct), with an AUC of 0.78. The model correctly identified 13/16 ACT cases (81%) and 14/20 higher-grade chondrosarcomas (70%). The weighted average F-score in the external cohort was 0.751 and the PRC AUC was 0.770. The Brier score was 0.25, reflecting moderate but not excellent calibration; the calibration curve confirmed the model's probability outputs were reasonable but had some deviation from perfect calibration.
Confusion matrix breakdown: In the training cohort, 9 ACT cases were misclassified as chondrosarcoma (false positives for CS) and 12 chondrosarcomas were misclassified as ACT (false negatives for CS). In the external test cohort, 3 ACT cases were misclassified as CS and 6 chondrosarcomas were misclassified as ACT. The asymmetry matters clinically: missing a high-grade chondrosarcoma (classifying it as ACT) is the more consequential error, as it could lead to undertreatment with curettage instead of resection.
Precision-recall analysis for chondrosarcoma detection: The precision-recall curve for G2-G4 chondrosarcoma detection in the external cohort showed an area of 0.857, suggesting the classifier maintained reasonable positive predictive value even at higher recall thresholds. For ACT, the PRC AUC was 0.661, reflecting the challenge of correctly flagging indolent lesions without over-calling malignancy. This trade-off between sensitivity and specificity for each class is clinically relevant when deciding the operating point for deployment.
The external cohort from centre 2 provided a unique opportunity for direct comparison because all 36 patients had both preoperative biopsy results and CT studies reviewed by an experienced musculoskeletal radiologist with 12 years of specialized bone sarcoma experience. The radiologist evaluated all CT studies blinded to tumour grade and disease course, classifying lesions based on four established imaging features: medullary cavity expansion with cortical thinning, cortical breakthrough, aggressive periosteal reaction, and presence of a soft-tissue mass. Any lesion showing at least one of these features was classified as higher-grade chondrosarcoma.
Biopsy performance: Preoperative biopsy had 64% accuracy (23/36 correct tumour grade), with an AUC of 0.66. Five biopsies returned inconclusive results, and 8 lesions were downgraded by biopsy (incorrectly classified as G1 when they were actually G2-G4). This downgrading led to 5 patients in centre 2 undergoing initial curettage followed by a second surgery for wide resection when pathology on the surgical specimen revealed higher-grade disease. The machine-learning classifier's accuracy was 75% vs. biopsy's 64%, a difference of 11 percentage points, though this did not reach statistical significance (p=0.29, McNemar's test). The authors note this could become significant in a larger cohort.
Radiologist performance: The experienced radiologist achieved 81% accuracy (29/36 correct), matching the machine-learning classifier's 75% without statistical difference (p=0.75). Looking at class-specific accuracy: the radiologist was 75% accurate for ACT (12/16 correct) and 85% accurate for higher-grade CS (17/20 correct). The radiologist's advantage in CS detection (85% vs. 70% for the classifier) likely reflects the clearer imaging signs that accompany more advanced disease, such as cortical breakthrough and soft-tissue mass, which are straightforward visual findings. For ACT, where imaging features are more subtle and overlap with benign enchondroma, both the radiologist and classifier struggled similarly.
Qualitative imaging findings: Among the 36 external test patients, cortical breakthrough was present in 16/20 CS cases (80%) and only 3/16 ACT cases (19%). Aggressive periostitis appeared in 14/20 CS (70%) but only 1/16 ACT (6%). Soft-tissue mass was exclusively found in CS (13/20, 65%). Maximum lesion diameter was substantially larger in CS (91 mm, IQR 59-124) vs. ACT (45 mm, IQR 31-54), confirming that size and aggressive imaging features correlate with grade but are not individually sufficient for reliable classification.
Retrospective design: Patient selection from electronic records spanning 10 years introduced selection bias, as only surgically treated patients with adequate imaging were eligible. Patients who were managed conservatively without surgery (watchful waiting for ACT) are entirely absent from this dataset. This is a real-world constraint for an uncommon disease, where prospective data collection would take many years to accumulate adequate numbers.
Bidimensional segmentation: While the authors argue that single-slice 2D segmentation matches or exceeds volumetric performance per recent literature, it captures only a cross-section of often-large, heterogeneous tumors. A G2 chondrosarcoma with focal dedifferentiation in a portion not captured by the selected axial slice could theoretically be misclassified if the chosen slice shows only lower-grade morphology. Volumetric segmentation might capture more of the tumour's spatial heterogeneity, albeit with greater time cost.
Interreader reliability assessed in only a subgroup: ICC for feature stability was measured in 30 of the 84 training patients, involving only 3 readers. This is the minimum requirement per Koo et al. guidelines but represents a relatively thin reliability dataset. Whether the selected features are equally stable across all radiologist skill levels, scanner types, or institutions beyond these two centres remains unknown.
Class imbalance and SMOTE: ACT outnumbered higher-grade CS roughly 2:1 in the training cohort, reflecting real-world epidemiology. SMOTE synthesizes new minority-class examples by interpolating between existing ones, which can introduce subtle artifacts if the minority class is small and highly variable. Although a 2:1 imbalance is generally considered manageable in ML studies, the synthetic CS examples may not perfectly represent the true diversity of G2-G4 chondrosarcoma morphology.
Different CT acquisition parameters across cohorts: The 64-slice scanner in centre 1 (1 mm slice thickness) and the 16-slice PET-CT in centre 2 (4 mm slice thickness, 500 mm FOV) differ substantially in spatial resolution and reconstruction parameters. While the authors resampled all images to 1x1 mm in-plane resolution, the coarser z-resolution in centre 2 means the single axial slices may capture different textural information. The fact that the model performed acceptably across both centres is encouraging, but scanner diversity should be expanded in future work.
Non-contrast CT only: Contrast-enhanced CT was not available for centre 2 patients, as PET-CT protocols at that institution did not routinely include contrast. Contrast enhancement could add information about tumour vascularity and matrix enhancement that correlates with grade. This limits the current pipeline to non-contrast acquisitions and leaves open the question of whether contrast-enhanced features would improve performance.
The study positions CT radiomics-based machine learning as a potential supplement to the existing multidisciplinary assessment framework, which currently combines clinical presentation, conventional imaging, and histology. The authors are explicit that the method is not proposed as a biopsy replacement; rather, it could serve as an additional non-invasive decision support layer, particularly useful in cases where biopsy is technically difficult, inconclusive, or carries elevated contamination risk. Integrating a radiomics probability score into the tumour board discussion is a plausible near-term application.
Contrast-enhanced CT radiomics: The absence of contrast enhancement is identified as a modifiable limitation. Contrast-enhanced CT could reveal differential enhancement kinetics between ACT (typically minimal enhancement) and higher-grade chondrosarcoma (more vascular, heterogeneous enhancement). Future studies incorporating these dynamic features could improve both sensitivity and calibration, and are encouraged by the authors as a logical next step.
Prospective validation: The retrospective design is the most important gap to close. A prospective multi-centre study collecting CT scans, radiomic features, biopsy results, and surgical pathology prospectively would provide a more rigorous test of real-world applicability. Given the rarity of chondrosarcoma, this will require consortium-level collaboration across multiple bone tumour centres, likely an international effort.
Context within prior radiomics work: The prior best in this space was an MRI radiomics study using an AdaBoostM1 classifier that achieved 75% accuracy in a single-institution hold-out test set of 58 patients. The present study matches that accuracy (75% on external validation) while using CT rather than MRI, doubling the population size, and importantly validating on an entirely independent external cohort rather than an internal hold-out. Both MRI and CT radiomics appear to carry comparable discriminative information, which opens the possibility of combining them for potentially higher accuracy. Feature-level fusion across imaging modalities has shown promise in other oncologic settings and warrants exploration here.
Broader applicability: The five features selected by the pipeline (Major Axis Length, NGTDM Contrast, GLRLM Short Run Low Gray Level Emphasis, GLCM Difference Entropy, GLCM Inverse Difference Moment) are not tumor-type specific and represent general texture and shape descriptors available in any PyRadiomics-compatible workflow. This means the pipeline infrastructure could be adapted with retraining for other cartilaginous lesion discrimination tasks, such as distinguishing enchondroma from ACT, or axial from appendicular chondrosarcoma, with minimal modification.