Central cartilage tumors of the long bones occupy a diagnostically difficult space in musculoskeletal oncology. The 2020 WHO classification separates these lesions into three categories that carry entirely different clinical trajectories: enchondroma (benign, no treatment needed), atypical cartilaginous tumor (ACT, locally aggressive with low metastatic potential, amenable to intralesional curettage or watchful waiting), and grade II chondrosarcoma (CS2, a malignant tumor requiring wide surgical resection with clear margins due to its metastatic potential). Misclassifying CS2 as ACT can result in inadequate surgery and disease recurrence; misclassifying ACT as CS2 can lead to unnecessary aggressive surgery with major functional consequences.
The incidence problem: Over the past three decades, the reported incidence of central cartilage tumors has risen substantially, driven mainly by incidental discovery on imaging performed for unrelated reasons. ACTs in particular are being detected more frequently than ever before. This growing case volume places increasing demand on musculoskeletal radiologists to accurately characterize these lesions, a task made harder by well-documented interobserver variability in lesion grading even among specialists at tertiary centers.
MRI radiomics as a solution, and its bottleneck: MRI-based radiomics combined with machine learning has emerged as one of the most promising techniques for differentiating ACT from CS2, with multiple studies reporting excellent diagnostic accuracy using T1-weighted sequences. However, the standard workflow in radiomics requires manual delineation of the region of interest (ROI) around the tumor, slice by slice, which is time-consuming, operator-dependent, and introduces its own variability. The reliability of downstream radiomics features depends heavily on the reproducibility of the input segmentation, making automated segmentation a critical unsolved problem in this workflow.
This study addresses exactly that gap. The authors propose and externally validate a deep learning model that can automatically delineate ACT and CS2 on T1-weighted MRI, with the stated goal of eventually combining automated segmentation with radiomics-based classification pipelines to create a fully automated, end-to-end diagnostic support tool for cartilaginous bone tumors.
This was a retrospective, Institutional Review Board-approved study conducted under the RADIO-BOSTT protocol (approved July 2022, Universita degli Studi di Milano). Informed consent was waived. The study enrolled patients from two tertiary bone tumor centers: center 1 at IRCCS Orthopedic Institute Galeazzi in Milan, Italy, and center 2 at Leiden University Medical Center in Leiden, Netherlands. The two-center design was essential for generating a genuinely independent external test cohort, which is a far stronger validity standard than internal cross-validation alone.
Cohort 1 (training, validation, and internal testing): This Italian cohort consisted of 99 consecutive patients with surgically treated and histologically confirmed ACT or CS2 of the long bones. Of these, 79 had ACT and 20 had CS2, reflecting the higher real-world incidence of ACT relative to CS2. Tumor locations included the femur (n=43), humerus (n=39), fibula (n=10), tibia (n=6), and radius (n=1). MRI scans were acquired on a 1.5-Tesla unit (Siemens Healthineers Magnetom Avanto or Espree).
Cohort 2 (external test only): The Dutch cohort comprised 65 patients, with 45 ACT and 20 CS2 distributed across the femur (n=46), humerus (n=10), and tibia (n=9). Imaging was performed on 3-Tesla or 1.5-Tesla Philips units (Ingenia or Intera), introducing genuine hardware variability relative to cohort 1. This cross-scanner, cross-country external validation is the key methodological strength of the study, as it tests whether the model generalizes beyond its training environment.
Inclusion criteria: Patients required a confirmed pathological diagnosis of central ACT or CS2 of the long bones after resection or curettage, plus an MRI performed within 3 months before surgery that included at least one T1-weighted sequence. Metacarpal, metatarsal, and phalangeal tumors were excluded, as were recurrent lesions and those with pathological fractures. The same two patient cohorts had previously been used in a companion study focused on radiomics-based machine learning classification, which provides context for the current work on automating the segmentation step that feeds into that classification workflow.
The deep learning architecture selected for this study is SEAGNET (Supervised Edge-Attention Guidance segmentation Network), originally proposed by Zhan et al. for malignant bone tumor segmentation, specifically osteosarcoma. SEAGNET uses a ResNet50 backbone for feature extraction, extended by a Feature Pyramid Network (FPN) and a custom decoder that incorporates mixed attention layers combining spatial and channel attention mechanisms, alongside dilation convolutions. The attention mechanisms allow the model to selectively emphasize diagnostically critical spatial regions and feature channels, which is particularly important for cartilage tumors where tumor boundaries against adjacent marrow and soft tissue can be subtle. The output layer uses a sigmoid activation function for binary segmentation (tumor vs. non-tumor), producing a single-channel prediction mask.
Ground truth segmentation: One musculoskeletal radiologist per center performed manual contour-focused segmentation using ITK-SNAP version 4.2.2. In cohort 1, the radiologist (A.C., 3 years of experience at a tertiary bone sarcoma center) drew ROIs enclosing the full tumor volume on all tumor-containing slices, including cartilage nodules separated by interspersed fatty marrow. In cohort 2, the radiologist (S.G., 5 years of experience) annotated only the single axial slice showing the largest tumor diameter. Noncontrast T1-weighted images were used throughout, based on prior evidence that T1-weighted radiomics alone can accurately classify cartilaginous tumors.
Training configuration: From the 99 patients in cohort 1, 1,037 tumor-containing slices were identified and split 70/15/15 into training (718 slices, 68 patients), validation (160 slices, 17 patients), and internal test (159 slices, 14 patients) sets. Images and ROIs were resized to 256x256 pixels. Pixel intensities were normalized per-slice to the 0-1 range using min-max normalization applied independently to each slice. The model was trained with the Adam optimizer at an initial learning rate of 1x10^-4, using a composite loss function combining binary cross-entropy and Dice loss to balance pixel-level accuracy with overlap-based similarity.
Data augmentation: To address the class imbalance (79 ACT vs. 20 CS2) and improve generalizability, data augmentation was applied to the training set, expanding tumor-containing slices from 718 to 3,590. Augmentation techniques included random horizontal and vertical flipping, rotation (10 to 70 degrees), width and height translations (pixel range -20 to 20), and zoom-in/zoom-out (range 0.8 to 1.2). For comparison, the study also evaluated nnU-Net, a widely used self-configuring segmentation framework, under the same conditions, with SEAGNET ultimately outperforming it in this population.
Model performance was quantified using two standard segmentation metrics: Dice Score (DS), which measures overlap between the predicted and ground truth ROI on a scale of 0 to 1 where 1 indicates perfect overlap, and Intersection over Union (IoU), the ratio of the intersection to the union of predicted and true segmentation areas, also scaled 0 to 1. Both metrics were computed in two ways: per patient (averaging across all slices for each patient, regardless of slice count) and per slice (treating each slice as an independent observation, giving patients with more tumor-containing slices proportionally more weight in the aggregate metric).
Internal validation set performance: In the validation subset of cohort 1, the model achieved per-patient DS of 0.782 +/- 0.148 and IoU of 0.663 +/- 0.175. Per-slice metrics were DS 0.742 +/- 0.273 and IoU 0.646 +/- 0.266. The higher standard deviation in per-slice metrics reflects the expected variability across individual cross-sections, where tumor boundaries differ in complexity from one slice to the next.
Internal test set performance: On the held-out internal test subset, per-patient DS was 0.748 +/- 0.191 overall (0.740 +/- 0.213 for ACT, 0.800 +/- 0.087 for CS2). Per-slice DS was 0.752 +/- 0.256 overall (0.746 +/- 0.268 for ACT, 0.771 +/- 0.220 for CS2). The slightly higher DS for CS2 compared to ACT in the internal test set may reflect the tendency of CS2 to form larger, more delineated masses compared to the sometimes more diffuse or lobular appearance of ACT.
External test set performance (primary result): The model's most clinically meaningful performance figure comes from the 65-patient external cohort at center 2, where one axial slice per patient was used. Here, the model achieved DS of 0.828 +/- 0.175 and IoU of 0.706 +/- 0.180. This external result actually surpasses the internal test performance, which the authors attribute to the fact that single representative slices at peak tumor diameter tend to be less challenging to segment than the full volumetric stack, which includes tumor margins and transition zones near the poles. Qualitative inspection of the predicted segmentations showed highly to moderately accurate delineations in both ACT and CS2, with illustrative examples showing individual cases achieving DS of 0.91-0.92 in well-segmented lesions.
One of the most practically informative findings in the study is the identification of a consistent failure mode: the model performed substantially worse on small lesions. Specifically, three patients in cohort 1 (one from validation, two from the internal test set) had Dice Scores below the threshold of 0.5, which is generally considered the minimum for clinically acceptable segmentation. All three of these cases had tumor volumes of 5.1 cm3 or smaller. In contrast, across the broader cohort, there was a visible positive association between tumor volume and segmentation accuracy, with larger lesions consistently yielding higher DS and IoU values.
Why small lesions are harder: In medical image segmentation, small target volumes present disproportionate challenges for convolutional neural networks. First, the ratio of tumor voxels to background voxels is extremely imbalanced, meaning the model's gradient signal during training is dominated by correctly predicting background pixels. Even the Dice loss component of the composite loss function, which partially mitigates class imbalance, does not fully solve this problem for very small objects. Second, with small lesions, even minor boundary errors represent a large fractional error in overlap metrics, causing DS and IoU values to drop sharply for what might be a visually acceptable segmentation.
Clinical relevance of this limitation: Small cartilaginous lesions are disproportionately likely to be ACTs rather than CS2 and are also the cases where watchful waiting is most commonly considered. Paradoxically, these are the cases where reliable automated segmentation for radiomics analysis would be most valuable, since they are also the most ambiguous diagnostically and the ones where the clinical decision between curettage and surveillance is least clear-cut. The model's lower performance on small lesions thus represents a meaningful gap between the current tool and what would be needed for full clinical deployment.
In the external test cohort (cohort 2), the one-slice-per-patient design may have partially masked this limitation, since the selected slice typically represents the largest cross-sectional area of the tumor, inherently avoiding the most challenging polar slices and very small lesions in the overall volume.
The authors evaluated nnU-Net as a direct comparator to SEAGNET in the supplementary material. nnU-Net is a self-configuring, automatic segmentation framework that has become a de facto benchmark in medical image segmentation tasks across oncology, consistently achieving top performance in public competitions for organ and tumor segmentation. Its architecture automatically adapts preprocessing, network topology, and training parameters based on dataset characteristics, which makes it a strong and fair baseline for comparison. Despite this, SEAGNET outperformed nnU-Net across the metrics reported in this study population.
Why SEAGNET has an architectural advantage here: The superiority of SEAGNET in this setting is attributed by the authors to its mixed attention mechanism, which simultaneously models spatial attention (which regions of the image are most relevant) and channel attention (which feature maps carry the most discriminative information). For cartilage tumor segmentation, where the lesion boundary with adjacent medullary fat is the critical feature, the spatial attention component helps the model focus on edge regions rather than the lesion interior. The Edge-Attention Guidance component explicitly encourages the network to learn edge-specific features during training, which is directly relevant to the contour-focused segmentation task.
Context from prior work: SEAGNET was originally proposed and validated for osteosarcoma segmentation on MRI, where it was shown to outperform classical architectures including U-Net, Fully Convolutional Networks (FCN), and DeepLab V3+. The current study extends this validation to cartilaginous tumors, a histologically and radiologically distinct entity from osteosarcoma but sharing the same musculoskeletal MRI imaging context. This consistency of SEAGNET outperforming alternatives across two different bone tumor types suggests the attention-based design is genuinely suited to this imaging domain.
The authors note that the performance gap between SEAGNET and nnU-Net may also reflect the relatively small size of the training dataset. Self-configuring frameworks like nnU-Net are optimized for robustness across datasets of varying sizes, but SEAGNET's explicit edge supervision may give it a more targeted advantage when training data is limited and the task is boundary-sensitive.
Class imbalance in both cohorts: ACT was substantially over-represented relative to CS2 in both center 1 (79 ACT vs. 20 CS2) and center 2 (45 ACT vs. 20 CS2), reflecting the true clinical incidence ratio. While data augmentation was applied to the training set to partially address this imbalance, the model still learned from a dataset where CS2 examples were roughly four times less common than ACT examples. This imbalance may limit the model's generalizability and robustness specifically for CS2 segmentation, which is the clinically higher-stakes classification. The authors acknowledge this but note the imbalance is inherent to the biology and epidemiology of these tumors rather than a sampling artifact.
Single-reader ground truth without interobserver reliability analysis: Each participating center used one musculoskeletal radiologist to generate the manual segmentation ground truth, and no formal interobserver reliability analysis was conducted as part of this study. Manual contour-focused segmentation has previously been demonstrated to be reproducible on both CT and MRI for central cartilage tumors in a separate publication, which the authors cite in support of their approach. However, the absence of dual-reader annotation with interobserver agreement metrics (such as intraclass correlation coefficients or Dice Score between readers) leaves some uncertainty about the upper bound of achievable automated performance, since the "ceiling" set by human agreement is not quantified here.
Asymmetric annotation depth between cohorts: Cohort 1 used full volumetric annotation (all tumor-containing slices), while cohort 2 used single-slice annotation at the plane of maximum diameter. This means the external validation measured two-dimensional segmentation accuracy at the most representative slice, rather than three-dimensional volumetric overlap. Volumetric accuracy across all slices, which is what would be needed for full radiomics feature extraction, was assessed only internally. The external test result (DS 0.828) should therefore be interpreted as a best-case estimate of single-slice performance, not necessarily as a proxy for full volumetric accuracy at an independent site.
Retrospective single-specialty design: The study is retrospective and restricted to patients who underwent surgery at two orthopedic oncology centers. Patients managed conservatively, including the growing population with ACT under watchful waiting, are entirely absent from the dataset. The performance of the model on incidentally discovered cartilaginous lesions in community radiology settings, which often have lower image quality, less standardized protocols, and a different case mix, has not been assessed.
The central stated goal of this research program is to combine automated deep learning segmentation with radiomics-based machine learning classifiers into a fully automated pipeline for ACT versus CS2 differentiation. The companion study by the same group already demonstrated excellent MRI radiomics accuracy for this classification task using manually drawn ROIs on T1-weighted images. The current work now provides the automated segmentation module that could replace manual annotation in that pipeline. Future work should directly test whether radiomic features extracted from automated SEAGNET-predicted ROIs yield classification accuracy comparable to features extracted from expert manual ROIs, since feature reproducibility and stability are known to differ between manual and automated segmentations.
Volumetric radiomics and 3D segmentation: The current study uses 2D slice-based segmentation as its operational mode, with the training and internal testing in cohort 1 performed on individual axial slices rather than full 3D tumor volumes. Extending the model to true 3D volumetric segmentation would enable extraction of 3D radiomic features, which have been shown in several musculoskeletal studies to offer reproducibility advantages over 2D features and to capture volumetric shape and heterogeneity information unavailable from single-slice analysis. Architectures such as 3D U-Net or adapting SEAGNET's attention mechanisms to 3D convolutions would be natural next steps.
Expanding the training dataset: With 99 patients in the training cohort, the dataset is relatively small by deep learning standards, even after augmentation. Federated learning approaches or multi-institutional data consortia would allow model training across larger, more diverse datasets without requiring direct data sharing. This is particularly relevant for rare bone tumors, where no single institution accumulates sufficient cases to train highly generalizable models. Larger training sets would also improve performance specifically for CS2 and for the small-lesion failure cases identified in the current study.
Clinical integration and prospective validation: Deployment in clinical practice would require prospective validation on consecutive MRI studies from radiological settings beyond tertiary bone tumor centers, including community hospitals and mixed-indication MRI scanners where cartilage tumors are frequently discovered incidentally. Integration into PACS or radiomics software platforms as a push-button tool, combined with output visualization that overlays the predicted segmentation on the original MRI for radiologist review and correction, is the practical endpoint. Regulatory pathways for such a tool would require demonstrating that radiologist-corrected automated segmentation yields equivalent or superior radiomics classification accuracy to radiologist-only manual segmentation at lower total time cost.