Author: Mustafa Bashir, Diana Kadi, Kyle J. Lafata, Jacob A. Macdonald, Mark Martin, Yuqi Wang, Marilyn Yamamoto 👨🔬
Affiliation: Duke University, Department of Radiation Oncology, Duke University, Department of Electrical and Computer Engineering, Duke University, Department of Radiology, Duke Unversity 🌍
Purpose: To develop a high-throughput, automated-data-interrogation pipeline for integrating imaging and clinical information to identify key determinants of liver volume (LV), enabling population-scale analysis and downsteam biomarker applications.
Methods: This retrospective study included patients who underwent abdominal CT or MR imaging within a single health system(2014-August 2024). The cohort was divided into a discovery set (2014-2021), a reserved testset(2022), and a future validation set (2023-August 2024) with new patient data (patients unseen in previous years). Images were processed using a high-throughput automated-computational pipeline featuring a trained nnUNet for liver segmentation utilizing parallel processing on 10 Nvidia RTX A6000 GPUs, and a ResNet-based quality-controller to exclude scans with >5% missing volume at liver edges. EHR data—including diagnosis codes, demographic variables, and physiological parameters—were linked to imaging within three months of scans. Predictors were selected using LASSO with cross-validation. Model performance was evaluated using weighted root mean square error(RMSE) and R². Nomograms were developed using linear regression on a core feature set identified via an elbow graph and a deep neural network(DNN) was trained to predict upper and lower bounds of LV to account for clinical variability and noise using LASSO selected features.
Results: We identified 78,983 unique patients with 145,165 total imaging series. Seven key predictors of LV were identified: body surface area, weight, age, liver steatosis, race, smoking status, and fever. Predictive models achieved RMSEs of 373.91–413.76 mL and R² of 0.41–0.44 across datasets. Incorporating diagnostic codes and lab results enhanced predictive performance and clinical utility. The DNN provided a reliable range of LV estimates, achieving an 88.43% coverage rate with an average range width of 5.95 ± 3.39 (z-score).
Conclusion: This study established an automated pipeline for integrating population-scale multimodal data. The approach highlights key determinants of LV and offers a framework for scalable, clinically relevant liver volumetry analysis.