CT image–metadata fusion predicts body weight from a single slice with 4.06 kg error

A deep-learning model fusing computed tomography (CT) image and metadata predicted body weight from one slice with mean absolute error 4.06 kg (95% confidence interval 3.90–4.23) and R²=0.92 in 3,479 held-out exams.

• Developed and compared three approaches on 30,857 training/validation exams from 16,169 patients using patient-disjoint 5-fold cross-validation, with a held-out test set of 3,479 exams from 1,796 patients.
• The fusion ResNet-18 achieved 69.5% of predictions within ±5 kg and 93.8% within ±10 kg; a weighted ensemble reduced mean absolute error to 3.99 kg, a 0.07 kg difference the authors deemed below practical significance.
• Split-conformal 90% and 95% prediction intervals showed empirical test coverage of 89% and 94% for case-level uncertainty quantification.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.