157 Automating detection and delineation of pancreas from surrounding tissue in pancreas retrieval photographs

Abstract Introduction Successful pancreas transplantation requires accurate assessment of donor pancreatic steatosis. Unlike other organs, the pancreas cannot be biopsied if intended for transplantation. Subjective visual assessment of steatosis is current standard practice. Artificial intelligence (AI) methods offer an objective alternative, but only if computers can first identify the organ from surrounding tissue (image segmentation). We present the first automated segmentation model designed to identify the pancreas from adjacent tissue. Methods Two independent image datasets of pancreata on the backbench were used. The first dataset, comprising 264 images, was split 75:25 for training and testing. The second dataset contained 58 images for external validation. Ground truth photograph annotations were verified by a transplant surgeon. We fine-tuned two open-source pre-trained segmentation models, Detectron2 and YOLOv8. Results Compared to human segmentation, in internal validation Detectron2 and YOLOv8 achieved accuracies of 98%, Dice coefficients (DIC) of 57% and 49%, and Areas Under the Receiver Operating Characteristic Curves (AUC) of 88% and 79%, respectively. In external validation, Detectron2 and YOLOv8 recorded accuracies of 98%, DIC of 70% and 47%, and AUC of 89% and 77%, respectively. Human annotation took a median of 33 (IQR:27-39) seconds. Detectron2 was 33 times faster at 0.99 (0.95-1.02) seconds, while YOLOv8 was 330 times faster at 0.09 (0.09-0.11) seconds. Conclusions The pancreas' resemblance to adjacent tissues challenges automatic segmentation. These models demonstrate high accuracy but lower DIC and AUROC, suggesting strong rule-out capabilities. Rule-in capabilities require further refinement with larger training sets, or more advanced segmentation pre-trained models.