Ensemble deep learning model based on CT scans: differentiating and subtype-classifying pancreatic inflammations and tumors, and predicting pancreatic lesion invasiveness.

Background: Pancreatic diseases, including pancreatitis and tumors, are often difficult to distinguish on imaging due to overlapping morphological features. The accurate classification of pancreatic lesions and preoperative prediction of invasiveness are critical for clinical decision-making and prognostic evaluation. The objective of this study was to develop an ensemble deep learning (DL) model based on computed tomography (CT) images for differentiating and subtyping pancreatic inflammations and tumors, as well as for predicting lesion invasiveness. Methods: This multi-center research included 6,740 patients' pancreatic CT images. An ensemble DL model integrating DeepLabV3, nnUNet-MS, and Adaptive Pyramidal Shifted Window-Swin-Transformer, respectively responsible for pancreatic segmentation, lesion segmentation, and diagnosis, was developed. Segmentation performance was evaluated using Dice and Intersection over Union (IoU); lesion differentiating and sub-classifying performance were assessed using sensitivity and accuracy. Gradient-weighted Class Activation Mapping (Grad-CAM) was conducted for conservatively treated pancreatic ductal adenocarcinoma (PDAC) to predict 12-month invasiveness, with imaging follow-up as reference standard. Results: The DeepLabV3 module had good pancreas segmentation performance (internal test set: Dice: 0.983; IoU: 0.971; external test set Dice: 0.981; IoU: 0.969). The nnUNet-MS module demonstrated superior lesion segmentation performance (internal validation set: Dice 0.941, IoU 0.932; external test set: Dice 0.942, IoU 0.930) to five other open-source nnUNet algorithms. The ensemble DL model showed high accuracy in both differentiating inflammatory and tumor lesions (internally 95.1%, externally 95.8%), as well as sub-classifying five inflammation subtypes and six tumor subtypes (internally 88.0% and externally 87.5%). It exhibited high sensitivity for detecting PDAC [93.9% (95% CI: 91.0-96.0%) internally and 92.9% (95% CI: 90.1-95.1%) externally]. In terms of predicting 12-month PDAC invasiveness, the volume difference between predicted and actual tumor progression ranged from 0.09 to 0.38 cm Conclusions: The integrated DL model demonstrated excellent performance in differentiating and subtype classifying pancreatic inflammation and tumor lesions, and in predicting the invasiveness of pancreatic lesions.