Performance Analysis of Pancreas Segmentation from CT Scan Standard Dataset and Synthetic Images Using Deep Learning

CT scan is the most popular technology used for diagnosis of pancreatic diseases. Due to the sedentary lifestyle the pancreatic diseases are rising in the population. Chances of developing pancreatic tumors are more in case of patients having pancreatic diseases like chronic pancreatitis. As the death rate after diagnosis is high in case of pancreatic cancer, the surgical procedures are very critical and should be planned strategically. This research work provides comparative analysis of some popular deep learning approaches used for segmentation of pancreatic region in CT scan images to facilitate early detection of pancreatic diseases. Two Standard public datasets having 19,723 annotated images data along with selected 1, 00, 268 DCGAN generated synthetic CT scan images were used for training, testing and validation. The performance of the deep learning models U-Net, Grid RCNN in combination with U-Net and hybrid CTO model was evaluated. Performance parameters like accuracy, Dice score, precision, recall and Jaccard Index were used for evaluation. CTO hybrid model including UNET along with Vision Transformer and Operator shows superior segmentation performance for the individual dataset as well as for cross validation. The proposed approach can be utilized in early detection and prognosis of pancreatic abnormalities.