Osteosarcoma is the most common primary malignant bone tumor in children and adolescents, and imaging plays an important role throughout diagnosis, staging, treatment planning, response assessment, and follow-up. Radiographs are often the first examination, while MRI and CT provide additional information on local tumor extent, bone destruction, and pulmonary metastases. However, subtle early findings, overlap with other bone diseases, and treatment-related changes can make image interpretation challenging. A review published in Medicine Plus examines the current evidence for AI and radiomics in osteosarcoma imaging. Researchers from Peking University People’s Hospital organized the evidence by clinical tasks, including radiographic triage, tumor segmentation, treatment response assessment, and risk prediction, while distinguishing osteosarcoma-specific studies from broader bone tumor studies used mainly as contextual evidence.
The review cautions that model performance should not be judged by reported metrics alone. Results can be influenced by patient numbers, study design, imaging protocols, and validation methods, and high scores on performance metrics such as AUC, Dice, or concordance index do not necessarily mean that a model will perform reliably in another hospital or patient population. Independent external testing is therefore important when assessing whether a model may generalize beyond the original study setting. Among current applications, radiograph-based artificial intelligence appears relatively close to clinical workflow integration. Some models have been developed to detect and classify suspicious lesions on full-field radiographs, and several studies have evaluated whether AI can assist radiologists. The review suggests that the most realistic near-term role is to support triage and referral decisions, such as prompting MRI, specialist referral, or biopsy planning, rather than providing an autonomous diagnosis.
Automated segmentation on CT and MRI is another active area of research. These methods may reduce manual contouring and support more consistent quantitative analysis, but many osteosarcoma-specific studies still include relatively small numbers of independent patients. The authors note that thousands of image slices do not represent thousands of independent clinical cases, and that a favorable overall Dice score may still conceal important local errors near joint surfaces, neurovascular structures, or small skip lesions. The use of AI to prediction of chemotherapy response, recurrence, metastasis, and survival remains more exploratory. Studies differ in imaging timing, sequences, clinical endpoints, and validation strategies, making direct comparison difficult. In this setting, reproducible performance in an independent external cohort may provide stronger evidence of clinical relevance than a higher score obtained only from internal testing.
Moving from promising algorithms to routine care will require progress across several areas at the same time. Osteosarcoma is rare and heterogeneous, making representative datasets difficult to assemble, while annotation and reference standards remain variable. Patient-level data separation, external validation, calibration, subgroup analysis, and failure-case review are not yet consistently reported, and multicenter studies must also account for differences in scanners, imaging protocols, and clinical definitions. Explainability methods may help clinicians review model behavior, but they do not by themselves establish biological or clinical validity. If such systems are eventually introduced into practice, changes in equipment, protocols, or patient populations could also affect performance, making prospective evaluation and ongoing monitoring important.
Overall, the review supports artificial intelligence as a potential decision-support tool rather than a replacement for radiologists or multidisciplinary teams. Future studies will need to show that these methods are reproducible across centers, can be integrated safely into clinical workflows, and provide measurable clinical value.
Medicine Plus
Literature review