An artificial intelligence-powered magnetic resonance imaging radiomic model predicted disease-free survival in patients with osteosarcoma. According to a recent study, the model demonstrated an accuracy exceeding 90% and strong biological interpretability by correlating imaging features with tumor immune biomarkers. It was developed and validated across 14 tertiary centers.
The model also achieved high predictive performance, with time-dependent areas under the curve of 0.916 in the training set, 0.802 in external test set 1, and 0.895 in external test set 2. Its sensitivity was approximately 92% and specificity was about 71% in the training set, plus approximately 80% sensitivity, 65% specificity, and 75% accuracy in test set 1 and 96% sensitivity, 60% specificity, and 81% accuracy in test set 2. The researchers further conducted a subgroup analysis on neoadjuvant chemotherapy (NAC), and found that the model's predictive accuracy was consistent independently from chemotherapy effects: area under the curve was 0.82 for NAC and 0.89 for non-NAC.
These results exceeded the accuracy of traditional prognostic factors such as tumor T stage and tumor volume, which yielded areas under the curve 0.63 or more. Kaplan–Meier analysis showed clear separation between high- and low-risk subgroups based on a radiomic threshold of 0.49, while calibration curves demonstrated close alignment between predicted and observed disease-free survival probabilities. Decision curve and clinical impact analyses confirmed superior clinical utility across probability thresholds of 65% to 100%.
The retrospective study, published in BMC Medicine, analyzed 270 patients with surgically treated, histologically confirmed osteosarcoma. The training set included 166 patients while 56 patients were in test set 1 and 48 patients were in test set 2. From contrast-enhanced fat-suppressed T1-weighted magnetic resonance imaging scans, 1,130 quantitative radiomic features were extracted and refined using dimensionality reduction and the Relief algorithm. Twelve key features—4 first-order and 8 textural—were incorporated into an Adaptive Boosting model optimized through 5-fold cross-validation.
To investigate biological interpretability, 41 hematoxylin and eosin–stained and 40 immunohistochemistry–stained whole slide images were analyzed. Ten nuclear morphological parameters were quantified to yield 150 patient-level features, and 5 tumor microenvironment biomarkers (CD3, CD8, CD68, FOXP3, and CAIX) were measured from immunohistochemistry slides. Correlation analyses highlighted that 9 of 12 radiomic first-order and textural features were significantly associated with 17 nuclear features and formed 32 radiology–pathology pairs—nearly 69% of which showed moderate correlations. Stronger relationships were observed between radiomic features and immune markers, including CD3, CD8, and CD8/FOXP3 ratio. “The varying degrees of correlations observed between H&E- and IHC-derived TME biomarkers and radiomic features validate the biological underpinnings of the radiomic features, affirming their interpretability at the cellular level,” noted lead author Qiuping Ren, PhD, of the Department of Radiology at The First Affiliated Hospital of Jinan University in Guangzhou, Guangdong, China, and colleagues.
The researchers found no significant correlations between radiomic features and CAIX, which they attributed to osteosarcoma's high vascularization and focal CAIX expression in necrotic zones often excluded during sampling, and suggested further validation in prospective settings and non-Chinese populations. Additional markers, such as HIF-1α, could be explored to better capture the impact of hypoxia on radiomic features, they added.
The researchers reported no conflicts of interest
Source: BMC Medicine