We evaluated the feasibility of using the Hounsfield unit (HU) value of the vertebral body to predict screw loosening in the thoracic spine. Consecutive patients who underwent thoracic spinal fusion surgery (from 2014 to 2020) were retrospectively identified. Patients with pedicle screw loosening in the upper instrumented vertebra (UIV) on postoperative computed tomography were included in the "loosening" group. The control group comprised an equal number of age-, sex-, and UIV-matched patients without screw loosening. Preoperative HU values at the UIV and lumbar T-scores were compared between the groups; receiver operating characteristic curves were constructed for HU values and T-scores to predict screw loosening, and the best cutoff values were determined. The same statistical analyses were performed for each subgroup, i.e., upper (T1-T4) and lower (T9-T12) thoracic levels. Forty-six patients each were included in the loosening and control groups. A significant between-group difference of HU values was noted for the lower thoracic UIV (loosening = 99.3, control = 126.3; p = 0.02) but not for the upper thoracic UIV (loosening = 171.8, control = 146.0, p = 0.70). T-scores did not differ between the groups for the lower (p = 0.14) and upper (p = 0.56) thoracic UIV. For the lower thoracic UIV, the area under the receiver operating characteristic curve was 0.660 for HUs (p = 0.01; 95% confidence interval [CI] 0.541-0.766) and 0.601 (p = 0.13; 95% CI 0.480-0.713) for T-scores. The optimal cutoff value for HUs was 126.3. Using this cutoff, HU values showed a better positive predictive value, negative predictive value, and accuracy compared to T-scores in predicting screw loosening.