Primary malignant bone tumors are rare, representing approximately 0.5% of all cancers worldwide, with an incidence of roughly 0.9 per 100,000 person-years in the United States. Osteosarcoma, Ewing sarcoma, and chondrosarcoma together account for about 80% of cases. Osteosarcoma has a pronounced peak incidence in adolescents and young adults, with about two-thirds of diagnoses occurring before age 25. The stakes are steep: five-year survival is around 75% for localized disease but drops below 25% once metastases are present.
The surgical and radiotherapy challenge: Wide surgical excision combined with multi-agent chemotherapy is the standard for high-grade osteosarcoma, but approximately 30% of pelvic cases are unresectable or only resectable with major functional compromise. In those patients, radiotherapy becomes the primary local treatment modality. Real-world utilization remains low, however. A SEER database analysis of 3,566 osteosarcoma patients found that only 11% received radiotherapy, and those who did had worse overall survival, largely reflecting selection of the most anatomically challenging, inoperable, and metastatic presentations. This context drives research into dose-escalation technologies that can achieve curative intent without unacceptable toxicity.
Scope of the review: This 2025 narrative review published in Frontiers in Oncology, authored by Tong, Chen, Li, Chen, and Yu, synthesizes the convergence of two distinct streams of innovation: advanced radiation delivery hardware (proton and carbon-ion therapy, MR-guided linear accelerators, and ultra-high-dose-rate FLASH radiotherapy) and artificial intelligence tools that accelerate and improve every step of the treatment planning workflow. The paper covers AI applications in imaging diagnosis, auto-contouring, knowledge-based planning, online adaptive replanning, quality assurance, and federated outcome modeling.
A core premise of the review is that these two streams reinforce each other. Hardware innovations create new degrees of freedom in dose delivery, while AI tools are needed to manage the resulting complexity in real time. The combination promises treatments that are individually tailored to tumor biology, patient anatomy, and daily physiological variation rather than standardized to population averages.
The review opens its technical survey with particle therapy, which has matured from single-center feasibility into multi-institutional phase II studies. The landmark trial cited enrolled 94 children and young adults with non-metastatic soft-tissue or bone sarcomas across multiple U.S. centers and delivered a median dose of 70 Gy (relative biological effectiveness, RBE) using proton therapy. After eight years of follow-up, local control and overall survival were both 77.5%, with grade 3 or higher late toxicity in fewer than 5% of patients. For skull-base lesions where standard photon dose limits are routinely exceeded by tumor proximity to critical structures, a 2023 evidence-based review of 14 series found that carbon-ion radiotherapy (CIRT) achieved approximately 80% five-year local control in chondrosarcoma while keeping severe early and late toxicities at or below 4% of patients.
Stereotactic body radiotherapy (SBRT): For oligometastatic or unresectable spinal and pelvic disease, dose-intensified SBRT has become standard. A 2023 systematic review pooling 1,137 spinal SBRT courses documented an overall pain-response rate of 83% (95% CI 68 to 94%) and durable control with single-fraction doses of 20 Gy or more. Contemporary vertebral-metastasis cohorts using 20 to 24 Gy in a single fraction report pain relief in 80 to 90% of patients at three months, with vertebral-compression fracture rates below 2% when strict dose constraints are adhered to.
MR-guided linear accelerators: High-field 1.5 T MR-Linac systems combine real-time volumetric imaging with on-table plan re-optimization. A 2024 systematic review of 26 prospective MR-guided studies concluded that online adaptation permits approximately 30% reduction in planning target volume (PTV) margins compared with CT-guided workflows, without sacrificing target coverage or extending beam-on time. In pelvic sarcoma series, median margin shrinkage from 10 mm to 7 mm has been achieved using deformable image registration and GPU-based re-optimization running in approximately 25 seconds.
FLASH radiotherapy: Ultra-high dose-rate delivery at 40 Gy per second or faster introduces a millisecond time dimension that pre-clinical evidence links to reduced oxygen depletion in normal tissues. The first-in-human FAST-01 phase I trial demonstrated workflow feasibility and durable pain palliation for extremity bone metastases using proton FLASH, with no grade 2 or higher acute toxicities at six months. Parallel murine models confirm equivalent tumor control alongside greater than 50% reduction in hematopoietic suppression compared with conventional dose rates, suggesting a clinically meaningful marrow-sparing effect directly relevant to pediatric patients receiving treatment near the spine or pelvis.
Distinguishing malignancy from infection on plain radiographs is notoriously difficult in musculoskeletal oncology, and delays in diagnosis directly worsen prognosis for sarcoma patients. The review highlights a multicentre study by Wang et al. that trained an ensemble of convolutional neural networks and Transformer models on 1,992 radiographs paired with clinical features. External test set performance was excellent, reaching an AUC of 0.963 and accuracy of 0.895, which matched senior radiologists and clearly surpassed junior readers. Crucially, saliency maps generated by the models localised cortical destruction and periosteal reaction patterns, providing the explainability that is necessary for adoption in multidisciplinary tumor boards.
Transformer-based early detection: A separate line of work specifically addresses early-stage osteosarcoma, where radiograph findings can be subtle. Transformer-based classifiers trained on annotated radiograph datasets achieve AUC above 0.90 for early osteosarcoma detection. The authors note that a novel three-class annotation method, distinguishing normal bone, benign lesion, and malignant lesion rather than the traditional binary approach, substantially improves model sensitivity for early-stage disease and reduces the false-negative rate that is most costly in this young patient population.
CT-based sarcoma subtype differentiation: Beyond detection and staging, AI models are being applied to subtype classification on CT images. One study trained multiple CT-based machine learning networks to differentiate pelvic and sacral osteosarcoma from Ewing sarcoma, tasks relevant to treatment selection because chemotherapy protocols differ significantly between these entities. Hybrid convolutional and Transformer architectures are now being re-trained on low-field MRI and dual-energy CT data, with the longer-term goal of building fully automated Enneking or AJCC staging dashboards that integrate imaging findings with clinical and laboratory variables to generate structured staging outputs at the point of acquisition.
The review frames AI-assisted diagnosis as particularly valuable in the sarcoma context because specialized musculoskeletal oncology expertise is concentrated in high-volume academic centers. AI tools could support earlier, more accurate referral decisions at community hospitals and radiology practices, compressing the diagnostic interval that currently averages several months for primary bone tumors in many healthcare systems.
Target delineation for osteogenic sarcomas is labor-intensive and prone to interobserver variability, particularly when skip lesions, post-biopsy artifact, and proximity to neurovascular structures complicate margin definition. The review centers on a study by Yin et al. that developed an nnU-Net-based segmentation pipeline trained on pelvic and extremity MRI data. The model achieved a mean Dice similarity coefficient (DSC) of 0.77 plus or minus 0.05 for gross tumor volume (GTV) segmentation, outperforming atlas-only workflows by 11 percentage points and cutting manual editing time by one-half. Although the test cohort was limited to 52 patients, the study incorporated cross-scanner data from multiple MRI systems and publicly released model weights to accelerate external validation.
Hybrid CNN-atlas cascades: Current developmental work focuses on hybrid architectures that combine atlas-based geometric priors with convolutional neural network-based edge refinement. The atlas propagates bony structural masks that constrain the CNN's search space, improving DSC at the tumor-marrow interface. This interface is clinically critical because local relapse in osteosarcoma most commonly arises from inadequate coverage of intramedullary disease extension, which is poorly visible on CT but detectable on MRI fluid-sensitive sequences.
Quantitative AI contouring benefits: Across the sarcoma literature surveyed in the review, deep-learning contouring tools consistently boost gross tumor volume Dice scores by 0.10 to 0.15 compared with atlas-only or manual workflows and cut manual editing time by approximately half. These gains translate directly into faster treatment planning cycles and more reproducible target volumes across treatment teams, both of which are prerequisites for the adaptive radiotherapy paradigms discussed elsewhere in the review.
The review acknowledges that contouring performance for organs at risk (OARs) in the pelvis and extremities has received less systematic evaluation in bone sarcoma specifically, with most published benchmarks coming from prostate, cervical, or head-and-neck cohorts. Transferring OAR segmentation tools to the heterogeneous bone-sarcoma anatomical landscape, which includes unusual post-surgical anatomy, metallic implants causing MRI artifact, and variable patient positioning, remains an active research frontier.
Knowledge-based planning (KBP) has evolved from dose-volume histogram templates into voxel-level, patient-specific dose prediction. The review's largest prospective KBP reference is a study by Cao et al. using a 3D U-Net residual architecture to predict volumetric modulated arc therapy (VMAT) dose distributions for 93 lung cases. Inference for a new patient took approximately 25 seconds on a single GPU, and 95% of re-plans generated from the AI-predicted dose distributions passed blinded physician review without any manual parameter tuning. The network has since been fine-tuned on high-grade sarcoma plans, where it reduces mean femoral head dose by 6 Gy while maintaining target dose conformity, a clinically significant reduction given the risk of femoral head necrosis in pelvic sarcoma patients.
Federated KBP for small-volume centers: Because high-grade sarcoma is rare, any single center's plan library is insufficient to train a robust dose-prediction model. The review highlights efforts to federate KBP models via secure containerized deployments: small sarcoma centers can benefit from the planning experience of high-volume institutes without exporting DICOM data. This model-sharing approach is an early practical implementation of federated learning for radiotherapy planning.
Online adaptive re-optimization: Daily anatomical change is particularly relevant for long-bone tumors receiving intensity-modulated proton boosts or for sacral chordoma treated on MR-Linac platforms, where tumor and organ positions can shift substantially between fractions. A convolutional LSTM sequence-to-sequence framework by Lee et al. predicts weekly tumor and esophagus geometry and auto-generates adapted plans within under one minute of cone-beam CT acquisition, achieving DSC above 0.75 across six treatment weeks and sparing mean esophageal dose by 4 Gy in validation patients. Pilot adoption of a similar reinforcement-learning optimizer on an MR-guided C-arm linac for pelvic sarcoma demonstrates plan-quality parity with manual re-planning in approximately 8 minutes total beam-hold time.
Together, these tools form a planning automation chain where dose prediction initializes optimizer parameters, the adaptive engine adjusts to daily anatomy, and final plan quality is checked by virtual QA before delivery, achieving a planning cycle that previously required hours within a clinically acceptable on-table time window.
Quality assurance (QA) is one of the most resource-intensive components of modern radiotherapy. Traditional patient-specific QA requires technicians to measure dose delivery for each plan on a phantom before the first treatment fraction. The UCSF "virtual QA" (VQA) system described in the review prospectively analyzed portal dosimetry signals and machine log-file metrics for 165 VMAT plans and identified 92% of plans that would have failed conventional measurement-based QA. This saved approximately 7 hours of technician time per week at a single center, and the approach is being extended to flag dosimetric errors before they enter the clinical queue rather than after plan approval.
Predictive machine performance monitoring: Beyond patient-specific QA, a machine learning model trained on daily "machine performance check" (MPC) log files predicted linear accelerator output drift 24 hours in advance with over 85% accuracy, enabling pre-emptive recalibration. This transforms QA from a reactive process of detecting failures after they occur to a predictive one that prevents unplanned downtime. Both tools are being integrated into commercial oncology information system dashboards, making AI-driven QA increasingly accessible to centers without dedicated medical physics informatics teams.
Federated survival modeling: Primary bone tumors are too rare for any single institution to power robust survival models. The review describes a seven-site Australian federated learning network that trained a Cox proportional hazards model for two-year overall survival on 1,655 non-small-cell lung cancer patients. The federated model achieved an AUC of 0.68 in prospective validation and maintained calibration when applied to a held-out 2017 to 2019 cohort, whereas locally trained models deteriorated to AUC at or below 0.63. Translational work adapting this infrastructure to bone sarcoma is underway, combining radiomics and circulating tumor DNA features. Preliminary cross-validation suggests a concordance index of approximately 0.75 versus 0.66 for single-center baselines, a meaningful improvement for a tumor type where the small dataset problem is most acute.
The authors argue that federated learning is not merely a technical convenience but a regulatory necessity for bone sarcoma AI, because the sample sizes required for external validation of late-toxicity nomograms, where follow-up spans decades, cannot realistically be achieved within a single institution's data governance framework.
The review transitions from technical description to clinical evidence by summarizing prospective series and registries that have begun to validate AI-augmented radiotherapy workflows in actual patients. A first-in-class series at MD Anderson treated four patients with deep-seated soft-tissue sarcomas on a 1.5 T Unity MR-Linac using an AI-assisted "adapt-to-shape-lite" workflow. After switching to AI guidance, median in-room time fell from approximately 90 minutes to 28 to 32 minutes per fraction, with residual setup error consistently below 1 mm. All four patients completed treatment without grade 3 or higher acute adverse events, and post-treatment imaging confirmed 100% target coverage and organ-at-risk sparing consistent with the original plan objectives.
MOMENTUM registry: Broader safety evidence comes from the international MOMENTUM registry, which by the time of publication had accumulated over 2,000 MR-Linac fractions across 9 countries. Acute grade 3 or higher toxicity occurred in only 1.4% of all patients and in just 0.4% of those treated with daily adaptive strategies, confirming that AI-guided online adaptation does not introduce new safety signals. These registry data are particularly important for bone sarcoma, where prospective randomized trials are difficult to conduct due to disease rarity.
Institut Curie pilot: At Institut Curie, the first four French sarcoma patients treated on an MR-Linac underwent daily AI-driven adaptive planning using an in-house contouring engine and 25-second GPU-based re-optimization. Median planning-plus-QA time per fraction was 30 minutes (range 27 to 35 minutes), and median intrafraction motion was 0.8 mm. No treatment interruptions or grade 2 or higher acute toxicities were observed. These numbers are significant because they demonstrate that a full AI-assisted adaptive workflow can compress on-table time to approximately half that of first-generation MR-guided approaches while maintaining sub-millimeter geometric accuracy.
Long-term proton outcomes: For pediatric and adolescent patients, where minimizing late effects is paramount, a phase II study from the Paul Scherrer Institute of pencil-beam scanning proton therapy for skull-base low-grade chondrosarcoma (n=77, median age 39) reached eight-year follow-up. Actuarial local control and overall survival at 8 years were 89.7% and 93.5%, respectively, with no grade 3 or higher cardiac or pulmonary toxicities. This eight-year safety benchmark aligns with a 2022 systematic review of 478 skull-base chordoma and chondrosarcoma cases where late grade 3 or higher toxicity remained below 5% across all organ systems.
Small, heterogeneous datasets: Primary bone tumors account for fewer than 1% of all malignancies, so most imaging AI studies have small cohorts with a median of approximately 112 patients and moderate to poor reporting quality. Cross-center variation in imaging protocols, including slice thickness and contrast timing differences, can cause models trained at single institutions to lose 10 to 20% of their external accuracy. Proposed mitigations such as self-supervised pre-training on unlabeled musculoskeletal scans and physics-guided data augmentation require prospective benchmarking before deployment claims can be substantiated.
Brittle external generalization: The review provides a sobering specific example: a deep-learning normal-tissue complication probability model for radiation-induced mandibular osteoradionecrosis achieved an AUC of 0.81 on its internal test set but dropped to 0.55 when tested at a different hospital, effectively reducing it to chance-level discrimination. This 26-percentage-point drop illustrates how internal cross-validation metrics routinely overestimate real-world utility. Published evidence for multicentre validation in bone sarcoma specifically remains almost entirely absent; the few networks with harmonisation layers or multimodal inputs have been validated on breast or lung datasets, not sarcoma.
Explainability and clinical acceptance: A survey of radiology explainability techniques cited in the review found that saliency maps, Grad-CAM, and SHAP values have become standard in published AI studies, yet fewer than 15% of models include prospective user-interface testing with actual clinicians. The consequences are visible in practice: in the French MR-Linac pilot, radiation oncologists overrode the AI-generated plan in 24% of fractions because contour boundaries could not be easily justified to the treating team. Research is shifting toward counterfactual explanations and interactive dose-volume trade-off dashboards; early prototypes reduce override rates by roughly one-third, but none have been validated specifically in bone sarcoma cohorts.
Regulatory and economic constraints: The EU AI Act (in force since August 2024) and China's Data Security Law classify clinical AI as high-risk, mandating third-party conformity assessment, post-market surveillance, and version control for adaptive algorithms. Economic viability of proton and carbon-ion centers as well as MR-Linac systems in low-resource settings remains uncertain given high upfront and maintenance costs alongside inconsistent reimbursement. The review calls explicitly for formal health-economic assessments to guide equitable access policy.
Multimodal foundation models ("Rad-Omics GPTs"): The review describes large-scale self-supervised vision foundation models as transformative for sarcoma AI. A Swin-Transformer framework called SMuRF fused CT voxels with whole-slide pathology data to predict HPV-related head-and-neck cancer outcomes, outperforming single-modality baselines by 12 percentage points in AUC. More directly applicable, ONCOPILOT is a promptable 3D CT foundation model trained on over 8,000 scans that produces 3D tumor masks in under one second with interactive editing capability, suggesting a single generalizable model could extend to sarcoma segmentation even with limited labeled training data. Coupling these vision encoders with radiogenomic Transformers ingesting RNA-seq panels could enable dose prescriptions dynamically adjusted to tumor molecular biology, an approach being explored in the PANORAMA trial for pelvic sarcoma (NCT05981234).
Digital-twin patients: A 2025 narrative meta-review catalogued over 40 oncology digital-twin prototypes that synchronize volumetric imaging, circulating tumor DNA, and electronic health records. Radiotherapy models were the fastest-growing subfield, generating day-by-day forecasts of local control probability and normal-tissue complication probability (NTCP). Mathematical oncology groups have demonstrated virtual dose-response curves that auto-adjust fractionation schedules when the simulated tumor control probability falls below a preset threshold, bringing what the authors term "anticipatory" radiotherapy into practical reach.
AI-optimized FLASH: Ultra-high dose-rate delivery cannot be safely managed by manual quality assurance alone, given the millisecond timescales involved. Bibliometric mapping shows a five-fold surge in FLASH-AI publications since 2021. Prototype amorphous-silicon detectors now track individual 2-microsecond micro-pulses, while deep-learning observers flag beam-current drift in real time, maintaining dose-rate variation within plus or minus 3%, a prerequisite for closed-loop adaptive FLASH therapy. Algorithms such as iDoTA predict full 3D photon or proton dose distributions in 50 to 100 milliseconds, enabling adaptive replanning within the same breath-hold.
Regulatory roadmap and international collaboration: The review proposes an international bone-sarcoma RT-AI alliance modeled on the MOMENTUM registry, which has already accumulated over 2,500 adaptive fractions across 40 sites with harmonized metadata. ESTRO 2025's dedicated session on foundation models in radiotherapy proposed shared ontologies for image-dose annotation to seed phase III AI-augmented sarcoma protocols. A three-horizon translational roadmap is offered: near-term AI auto-segmentation and KBP; mid-term digital twins and integrated radiomic-genomic stratification; and long-term randomized clinical validation and standardized FLASH adoption. Federated learning with formal differential-privacy budgets (epsilon below 1) has demonstrated less than 2% accuracy penalty in multi-country imaging tasks and is proposed as the default compliance strategy under existing and anticipated data governance laws.