The Radiation Oncology Feed For radiation oncologists · physicists · RTTs · dosimetrists
preliminary nsclc sbrt

Time-series CT radiomics predicted rib fracture after lung SBRT with 0.79 validation AUC

A longitudinal CT radiomics model identified post-SBRT rib fracture risk with 85% sensitivity and 71% specificity in validation.

RadOncToday · ·2 min read

KEY POINTS

  • This single-institution retrospective study included 67 patients with NSCLC, comprising 1,605 individual ribs, analyzed using CT scans at three timepoints: within one month before SBRT, 2–3 months after treatment, and 5–6 months after treatment. Median age was 71 years, median follow-up was 796 days, and 50 Gy in 5 fractions was the most common regimen, used for 62.2% of 74 treated lesions.
  • Rib fractures were ultimately observed in 22 of 67 patients (32.8%), but events were rare at the individual-rib level: 16 of 963 ribs (1.7%) in the training cohort and 13 of 642 ribs (2.0%) in the validation cohort fractured. No fractures had occurred by the final CT used as model input, meaning the model was attempting to identify changes preceding the clinical event.
  • The Knowledge-aware Temporal Mixture of Experts model integrated 1,051 radiomic features from each of three CT timepoints. Because only 16 fractured ribs were available for initial training, TimeGAN generated 837 synthetic fractured samples, producing a near-balanced training dataset of 1,800 instances.
  • KA-TMoE achieved a validation ROC-AUC of 0.792, compared with 0.577 for the same architecture without knowledge-aware gating. Performance also fell markedly when any one CT timepoint was removed, with AUCs of 0.561, 0.552, and 0.556, while single-timepoint models achieved only 0.533–0.553.
  • At the prespecified training-derived threshold of 0.391, the model correctly identified 11 of 13 fractured ribs in the validation cohort, giving 85% sensitivity and 71% specificity. The model also had the lowest Brier score at 0.0196, indicating the best calibration among the tested approaches.
  • Model-defined high-risk ribs fractured substantially earlier and more frequently than low-risk ribs, with a hazard ratio of 10.82 (p<0.001). In multivariable analysis, the KA-TMoE output remained independently associated with fracture in validation with an odds ratio of 12.05 (p=0.002), while the evaluated dosimetric variables were not significantly associated with fracture risk.
  • Most of the strongest predictors were texture rather than shape features, and approximately half of the 20 most influential features came from the 5–6-month post-SBRT CT. A stricter patient-level split still produced an AUC of 0.761, but performance declined substantially without synthetic augmentation, reinforcing the need for larger external datasets before clinical use.

CLINICAL TAKEAWAY

Routine surveillance CT may contain early textural signals of radiation-induced rib injury months before a fracture becomes visible. The concept is clinically interesting, but with only 29 fractured ribs overall, single-center imaging and substantial reliance on synthetic training data, this is a hypothesis-generating prediction model rather than a tool ready to influence SBRT planning or follow-up.

SOURCE

International Journal of Radiation Oncology, Biology, Physics