Single-slice ct body-weight prediction with image and metadata fusion: a multi-model comparison.

Authors

Abstract

A measurable fraction of examinations in retrospective imaging archives lack a recorded body weight, limiting metadata completeness for tasks such as cohort auditing, weight-stratified analysis, and the assembly of machine-learning training datasets. We developed and compared three approaches for body-weight prediction from a single middle-slice CT image: a LightGBM model on 19 CT-derived tabular features (water equivalent diameter, effective diameter, body composition, scanner parameters), an ImageNet-pretrained ResNet-18 on images alone, and a ResNet-18 with late fusion of the same 19 tabular features into the 512-dimensional image embedding. Models were trained with 5-fold patient-disjoint cross-validation on a combined train+validation development pool of 30,857 examinations from 16,169 patients; the best-performing fold of each model family was selected by within-fold validation mean absolute error (MAE) and evaluated, without further tuning, on a held-out test partition of 3,479 examinations from 1,796 patients (patient-disjoint from training and validation). On the test set, the fusion ResNet-18 achieved MAE 4.06 kg (95% CI 3.90-4.23; R² = 0.92), with 69.5% of predictions within ± 5 kg and 93.8% within ± 10 kg. A weighted ensemble of the fusion ResNet and LightGBM produced a small further reduction in MAE to 3.99 kg (paired bootstrap p < .001), but the 0.07 kg difference is below practically meaningful thresholds. The simpler single-model ResNet-18 with metadata fusion may be useful for retrospective missing-weight imputation, with split-conformal 90% and 95% prediction intervals (empirical test coverage 89% and 94%) available for case-level uncertainty quantification.

Tags

Topics