A new deep learning approach for pancreas segmentation on abdomen CT images: Pascal U-Net

Nowadays, the use of deep learning models in medical image processing has gained momentum. Especially in studies on organ segmentation from slice images, deep learning methods are frequently preferred. Since the pancreas, located in the abdominal region, differs in shape, location and size in each person, its segmentation is quite challenging. To solve this problem, the U-Net model, which is one of the deep learning models, is generally preferred in the literature. In this thesis, a new deep learning model based on the U-Net model with an architecture suitable for the number sequence in Pascal's triangle has been proposed for pancreatic segmentation. This proposed model is named Pascal U-Net model and the performance of the model is evaluated on two different data sets. First, The Cancer Imaging Archive Pancreas-CT dataset, which is a publicly available and frequently used dataset in the literature, was used. In addition, abdominal CT images taken from the Department of Radiology at Selcuk University Medical Faculty Hospital were used as the second data set. A slice image was selected for each patient from the records in the datasets and datasets for deep learning networks were created by applying preprocessing methods. In order to compare the pancreatic segmentation results obtained on both data sets with Pascal U-Net model, segmentation process was also performed on the same data sets with the U-Net model. Segmentation maps obtained as a result of 2, 4 and 6 fold cross validation and deep learning models run on different batch sizes from 1 to 10, were evaluated using 7 different performance metrics. Pancreas segmentation results performed with each batch size and different fold cross validation are the average of 10 run results. Both U-Net and Pascal U-Net segmentation results were analyzed based on 7 different metrics and visual evaluations. When the results are examined; in both data sets, Pascal U-Net model outperformed traditional U-Net architecture with a value of approximately 1% in terms of Dice Similarity Coefficient metric.