A Novel Deep Learning-Based Automated Adipose Tissue Segmentation Model for Value-Added Screening at Abdominal CT

Purpose: The increasing global obesity prevalence is concerning due to its link with numerous chronic diseases and premature death. The primary health assessment for obesity, body mass index (BMI), fails to consider adipose and muscle distribution. The visceral-to-subcutaneous ratio (VSR) of adipose tissue provides a more accurate measure of metabolic health than BMI but is not routinely used. VSR can be obtained from abdominal CT scans, which are often underutilized for this purpose. Manual segmentation of VSR and muscle is not scalable, but automatic segmentation models can analyze CT scans effectively. We propose a new deep learning-based model for adipose segmentation for faster processing time. Methods : A large, diverse dataset from University of Wisconsin CT scans comprising 430942 images from 8671 exams and 6482 unique patients, was split into training (80%) and validation (20%). The deep learning-based model (DLBM) was a U-Net variant with a VGG11 encoder, trained using the Adam optimizer, and tested against an established feature-based model (FBM) for automatic adipose segmentation. Four readers reviewed 3200 image pairs to compare the FBM and DLBM adipose segmentation and assess interrater reliability. Survival analysis, hazard modeling, and AUC analysis were also conducted. Results . Our sample included unique CT scans from 1723 patients with colorectal cancer with 881 women (51.13%) and 842 men (48.87%). The DLBM was 1000 times faster than the FBM in image processing and failed in 3 (0.2%) scans, while the FBM failed in 55 (3.1%) scans. Radiologist readers preferred the L3 adipose tissue segmentation by the deep-learning based model in 66.3% of cases over the feature-based model while the FBM was preferred in only 7.5%. Patients’ survival was 52.41% (n=903). Kaplan-Meier survival analysis and Cox proportional hazard ratio models between the FBM and DLDM VSR hazard ratios were nearly identical. Conclusion . The DLBM for adipose tissue segmentation was faster, failed less frequently, and was more often preferred by radiologist readers with similar predictive power for survival compared with the FBM. The DLBM enhances performance and segmentation speed without sacrificing predictive ability, potentially leading to faster patient record processing and computational savings when scaled up.