Deep learning-based body weight estimation from lower abdominal CT scans

Abstract Body weight is a key parameter for assessing the risk of an osteoporotic hip fracture by autonomous finite element analysis based on lower abdominal computed tomography (CT) scans. To allow opportunistic diagnosis of this fracture risk by any lower abdominal ? CT scan, we herein present a deep learning algorithm for estimating the patient’s body weight directly from lower abdominal CT scans. The algorithm is composed of two automatic steps: pre-processing of the CT scan followed by a two-branch convolutional neural network combining image features from axial CT slices with patient metadata. The algorithm was trained on 100 CT scans from different CT scanners and tested on 24 held-out test subjects. The body weight estimation has a mean absolute error (MAE) of 3.62 kg, and a root mean squared error (RMSE) of 4.26 kg on the 24 test set, with 75.0% of predictions within ± 5 kg and 100% within ± 10 kg of actual weight. The Pearson correlation coefficient was 0.93 (R 2 = 0.88), demonstrating a good predictive performance. To further validate these results, a five-fold cross-validation was performed, yielding an average MAE of 7.43 ± 1.14 kg and an average RMSE of 9.65 ± 1.59 kg. On average across all folds, 45.0% of predictions were within ± 5 kg and 72% were within ± 10 kg of the actual weight. This automated algorithm for determining body weight allows an opportunistic diagnosis of osteoporotic hip fracture from lower abdominal CT scans.