CT radiomics tracks vertebral fragility after esophageal cancer radiotherapy
Thoracic vertebral compression fracture emerged as a measurable late toxicity signal in patients with esophageal squamous cell carcinoma treated with radiotherapy, and pretreatment CT carried more information than bone density alone. In this two-center retrospective cohort, investigators built and externally tested a radiomics-based prediction strategy meant to stratify individual fracture risk before treatment began.
The clinical question is straightforward, but the answer is not. Thoracic vertebrae sit inside the radiation field for many esophageal cancer plans, and repeated dose exposure can weaken already vulnerable bone. When fracture risk is missed, the consequence is not abstract: pain, reduced mobility, and a new source of morbidity layered onto an already difficult cancer course.
Cohort structure and event burden
The study included 551 patients with pathologically confirmed ESCC who underwent radiotherapy. Patients from institution 1 were split into a training cohort of 321 and an internal validation cohort of 138, while 92 patients from institution 2 formed the external validation cohort. Radiomic features and CT attenuation values were averaged across all eligible irradiated thoracic vertebrae on pretreatment CT for each patient, giving the model a vertebra-level imaging signature before the first treatment fraction.
Follow-up was short but clinically meaningful for an early toxicity endpoint. Median follow-up reached 0.92 years, with an interquartile range of 0.58 to 1.50 years, and 77 patients developed post-radiotherapy thoracic vertebral compression fractures. Events were distributed across the cohorts as 47 in the training set, 21 in the internal validation set, and 9 in the external validation set. Those numbers matter because fracture prediction models can look strong in development and then thin out quickly when tested against smaller independent samples.
Age, alcohol consumption, radiotherapy fractions, vertebral CT attenuation, and Radscore survived selection into the final radiomics-integrated model. The retained predictors make biological sense together. Older age and alcohol exposure both track with skeletal vulnerability, radiotherapy fractions reflect cumulative tissue injury, and vertebral CT attenuation offers a structural proxy for bone quality. Radiomics adds a second layer by quantifying heterogeneity in texture and density patterns that conventional visual review cannot reliably capture.
Performance of the integrated model across three cohorts
The final radiomics-integrated model performed well in the development cohort and retained useful discrimination in both validation sets. The C-index reached 0.888 in training, 0.796 in internal validation, and 0.847 in external validation. At 24 months, the time-dependent AUC values were 0.903, 0.716, and 0.863, respectively. Bootstrap validation yielded an optimism-corrected C-index of 0.854, suggesting some shrinkage after internal correction but leaving the model still in a clinically respectable range.
One sentence captures the central result plainly: the model separated higher-risk from lower-risk patients better than attenuation alone. The comparison mattered most in the external cohort, where the C-index difference between Models 6 and 5 reached statistical significance with _p_ = 0.018.
Still, the incremental value was not uniformly confirmed by all reclassification metrics. The 24-month NRI was not significant in the external cohort (_p_ = 0.186), and the IDI also fell short of significance (_p_ = 0.067). Those mixed results point to a familiar problem in imaging biomarker work: a model can discriminate reasonably well and still fail to show a clearly superior reclassification benefit when event counts are limited. Calibration and decision-curve findings were therefore interpreted with caution, especially because the external cohort contained only nine fracture events.
Why attenuation alone is not enough
Vertebral CT attenuation is an appealing variable because it is already available on routine scans and reflects bone mineral status indirectly. In radiation oncology, however, the vertebra is not simply a passive structure measured once at baseline. It is exposed to a dose pattern shaped by tumor location, field design, fractionation schedule, and the patient’s underlying skeletal reserve. A plain attenuation value cannot fully capture microarchitectural irregularity, marrow composition, or subtle textural change across adjacent vertebral bodies.
Radiomics is designed to extract those hidden patterns. By quantifying gray-level relationships and dependence features inside the vertebral body, it can describe imaging complexity that may align with fragility before macroscopic collapse appears. In a setting where fracture prevention depends on knowing who is most vulnerable before treatment starts, that extra signal has obvious appeal. The model’s retention of both attenuation and Radscore suggests the two measures are complementary rather than redundant.
What the study can and cannot support yet
This work offers a plausible path toward individualized risk stratification, but not a finished clinical tool. The design was retrospective, and the external cohort was small enough that every event carried disproportionate statistical weight. A single-site quirk in imaging acquisition, segmentation practice, or radiotherapy planning could still move performance estimates, even with two centers involved.
The median follow-up under one year also limits how far the data can stretch into late skeletal toxicity. Vertebral compression fractures may continue to accrue beyond the observation window, so longer surveillance could change both the absolute event rate and the apparent calibration of the model. The investigators also evaluated several model structures, which is useful for development but raises the usual concern that the most favorable configuration may travel less smoothly into another hospital system.
Even with those caveats, the study places CT radiomics squarely in the conversation for radiotherapy planning in ESCC. If a patient’s pretreatment scan already contains a reproducible signal for vertebral fragility, clinicians could eventually use that information to intensify bone monitoring, refine dose planning, or identify patients who may need closer follow-up after treatment. The present data do not yet prove that such steps will reduce fractures, but they do show that the risk is visible on imaging before the injury occurs.
External validation still needs more fracture events
The final barrier is not whether the model can be built. It is whether it can be trusted across larger independent cohorts with enough fractures to test real-world utility, not just statistical separation. Future validation will need broader case volumes, consistent CT acquisition, and enough outcome events to judge whether the observed gain in discrimination translates into actionable clinical decisions at the bedside.