Deep Learning to Detect Body Composition and Its Role in Developing Postoperative Pancreatic Surgery Complications

Background

Variance in skeletal muscle area (SMA), visceral adipose tissue (VAT) and subcutaneous adipose tissue (SAT) negatively impacts outcomes after pancreas surgery. We aim to incorporate an existing deep learning algorithm automating body composition segmentation from computed tomography (CT) for accurate and rapid risk identification.

Methods

We conducted a retrospective study of patients having pancreatic surgery at a high-volume centre (2016–2021). Using a deep learning algorithm, we analysed preoperative CT images at the L3 level for SMA, VAT, SAT and IMAT (AutoMATiCA, Cambridge, MA, USA). Two board-certified radiologists validated the analysis. Skeletal muscle index (SMI), VAT and VAT/SAT ratio were calculated. We then evaluated the incidence of pancreas surgery-specific, pulmonary, noninfectious and infectious outcomes.

Results

We reviewed 158 patients: median (IQR) age 67.6 (61.6, 75.3) years; female (52.5%); pancreatic cancer diagnoses (65.8%); and Whipple procedure (81%). Automated body composition calculation time for all patients was 553 s. Patients experiencing composite sepsis complications had higher VAT (193.7 [IQR 132.7, 249.7] vs. 146.2 [IQR 87.3, 220.5], p = 0.029). Additionally, patients experiencing composite infectious complications had higher VAT (193.7 [IQR 133.4, 277.5] vs. 143.1 [IQR 72.2, 202.8], p = 0.041). VAT was also higher in patients with noninfectious complications (274.9 [IQR 228.0, 329.8] vs. 148.7 [IQR 90.9, 221.0]; p = 0.020). Other anthropomorphic features, such as SMA, SAT and IMAT, did not have any relation to postoperative composite outcomes.

Conclusions

Higher visceral adipose tissue was associated with worse outcomes after pancreas surgery. Deep learning applied to CT scans may be valuable for identifying at-risk body compositions associated with adverse surgical outcomes. Further studies are needed to confirm these findings.