Integrative deep learning analysis of 2D and 3D body composition features for predicting postoperative pancreatic fistula after distal pancreatectomy.

The incidence of clinically relevant postoperative pancreatic fistula (Grade B/C) was 47.4% in the study cohort.

• Study comprised 230 patients with pancreatic cancer from two institutions between 2015 and 2022
• Mean age of patients was 62 years
• POPF was defined as Grade B/C according to standard classification
• Retrospective analysis was conducted across two institutions

The deep learning segmentation model achieved high accuracy for abdominal muscle segmentation across different regions.

• Dice similarity coefficients (DSCs) for muscle segmentation in the testing set ranged between 91.84% and 98.41% across different regions of the abdominal musculature
• Segmentation was performed using preoperative CT images
• Both 2D and 3D segmentation approaches were developed and evaluated

The deep learning model achieved high segmentation accuracy for visceral and subcutaneous adipose tissue.

• DSC for visceral adipose tissue segmentation in the testing set was 97.10%
• DSC for subcutaneous adipose tissue segmentation in the testing set was 98.57%
• Segmentation was performed on preoperative CT images

The integrated clinical and imaging-based POPF prediction model using Gradient Boosting Decision Trees demonstrated superior predictive performance in the external test set.

• The model achieved an AUC of 0.82 in the external test set
• Sensitivity was 0.81 and specificity was 0.76 in the external test set
• Model performance was evaluated using both AUC and decision curve analysis (DCA)
• The predictive model combined both clinical and imaging-based body composition features
• Gradient Boosting Decision Trees (GBDT) were used as the predictive modeling framework

The study developed a combined 2D and 3D body composition analysis framework using preoperative CT imaging to assess abdominal muscles and fat.

• The pipeline analyzed both 2D and 3D features of body composition
• Abdominal muscles and fat (visceral and subcutaneous) were segmented using the deep learning model
• Preoperative CT images served as the input data source
• Performance was evaluated on both validation and external test datasets from two institutions