Comparative analysis of deep learning-based segmentation models on pancreatic images

The pancreas is one of the vital organs of the human body, responsible for breaking down dietary fats through the lipase enzyme and performing various hormonal functions. Its endocrine role involves producing hormones that regulate blood glucose levels. Insulin is perhaps the most significant example, and additionally, the pancreas facilitates normal hormonal secretion in certain parts of the digestive system. Serious health problems may arise when these two functions do not work in harmony. Pancreatic pathologies include cancer, diabetes, pancreatitis, hormonal disorders, and digestive complications. Since most pancreatic disorders are life-threatening, they are among the investigations prioritized by physicians. Initial detection of pancreatic diseases is performed using blood and urine tests. Additional imaging techniques such as CT (Computed Tomography), MRI (Magnetic Resonance Imaging), and EUS (Endoscopic Ultrasonography) are utilized for further examination if necessary. These methods allow detailed evaluation of the pancreas in terms of its structure and function. The identification and delineation of pancreatic boundaries on medical images are called "pancreatic segmentation." The segmentation process is essential for accurate pancreatic assessment. The pancreas can vary significantly in size and shape from one individual to another. Manual segmentation is time-consuming and does not yield uniform results across all physicians. This task requires experience and may be prolonged and labor-intensive. Providing a pre-segmented visual representation to clinicians during the assessment process may accelerate evaluation and help maintain longer concentration. In recent years, interest in deep learning-based organ segmentation methods has increased in order to address these challenges and to obtain more accurate results. Deep learning has proven to be a high-potential machine learning application that enhances precision in tasks such as diagnosis and classification of medical images with high accuracy. Existing segmentation techniques based on deep learning are widely used due to their improved consistency and speed compared to human effort. In this study, two separate experiments were conducted for pancreatic segmentation using a publicly available dataset. In the first experiment, U-Net and V-Net architectures were compared. The dataset was split into 80% for training and 20% for validation. The findings showed that the U-Net architecture achieved a validation accuracy of 99.78 with a validation loss of 0.041. The V-Net architecture achieved a validation accuracy of 99.75 and a validation loss of 0.046. In the second experiment, the same dataset was divided into 80% for training, 10% for validation, and 10% for testing. The U-Net, V-Net, Attention-based U-Net3B, TransU-Net3B, SwinU-Net3B, and FocalU-Net3B deep learning architectures were compared. The Dice score results were obtained as follows: U-Net 86.55, V-Net 82.09, Attention-based U-Net3B 85.47, TransU-Net3B 85.28, SwinU-Net3B 85.63, and FocalU-Net3B 84.55. The high accuracy rates of the architectures demonstrate the success of deep learning approaches in pancreatic segmentation. Pancreatic segmentation is extremely important, especially in the surgical context of pancreatic diseases. Surgery is considered one of the most effective treatments for life-threatening conditions such as pancreatic cancer. However, the successful completion of surgery requires accurate delineation of the pancreas, which is essential for both patients and clinicians. In addition to surgical intervention, segmentation is also critical for evaluating disease progression and analyzing treatment effectiveness. Therefore, the accuracy and speed of segmentation are crucial for both diagnosis and treatment. From the patient's perspective, segmentation plays a significant role in disease identification. Methodologies based on deep learning algorithms, which help minimize errors in these processes, enable earlier and more accurate diagnosis of diseases.