2D and 2.5D Pancreas and Tumor Segmentation in Heterogeneous CT Images of PDAC Patients.

Pancreatic cancer is a major contributor to global cancer mortality, emphasizing the critical importance of early detection for enhancing patient survival rates. However, its asymptomatic nature until advanced stages poses significant challenges. Computed tomography (CT) is widely used for the initial evaluation of cancer lesions, providing an important aid for diagnosis and therapy decisions. Yet, identification and segmentation of pancreatic lesions remain problematic due to tissue density confusion, morphological variations, and indistinct borders. Deep Convolutional Neural Networks (DCNNs) offer a promising solution, excelling at detecting subtle features imperceptible to the human eye. This paper presents initial findings from a study investigating the impact of data preparation and training methods on pancreatic tumor segmentation using DCNNs. Leveraging the Medical Segmentation Decathlon (MSD) dataset for training and the Clinical Proteomic Tumor Analysis Consortium-Pancreatic Ductal Adenocarcinoma (CPTAC-PDA) dataset for testing, diverse techniques including preprocessing, training configurations, and augmentation are evaluated. Notably, 2.5D models with augmentation exhibit significant performance improvements. Evaluation on the CPTAC-PDA dataset confirms the robustness of DCNN models, emphasizing their potential in enhancing pancreatic tumor segmentation for early intervention. These results underscore the critical role of deep learning in addressing the challenges of pancreatic cancer management, emphasizing the urgency of early diagnosis and intervention strategies.