Deep learning model for body weight estimation from computed tomography scout images incorporating sex and height

Accurate body weight measurement is essential for determining the appropriate dose of contrast agent in contrast-enhanced computed tomography (CT) examinations. However, in emergency medicine, obtaining accurate measurements is often challenging, which can lead to over- or underdosing of contrast medium. Therefore, we aimed to develop and evaluate a deep learning model that estimates body weight using chest-abdominal CT scout images, sex, and height. We retrospectively analyzed the data of 763 hospitalized patients whose CT examination dates matched their weight-measurement dates. This dataset included patients with arms positioned alongside the body and those with metallic implants, commonly encountered in emergency medicine. After performing five-fold cross-validation, a deep learning model based on transfer learning with VGG16 was constructed. The following four input combinations were evaluated: (1) scout images alone; (2) scout images with sex; (3) scout images with height; and (4) scout images with sex and height. The percentages of cases with differences between predicted and actual body weights within ± 5 kg were 84.3%, 90.2%, 92.8%, and 90.2% for inputs (1)-(4), respectively. The corresponding mean absolute percentage errors were 4.8%, 4.7%, 4.1%, and 4.0%, respectively. Our method provides a useful tool for estimating body weight in patients of unknown weight, with its accuracy appearing largely unaffected even when the patients had arms positioned alongside the body or possessed metallic implants. Moreover, incorporating sex and height into the scout images further improved the prediction accuracy.