Artificial Intelligence for Diagnosis of Pancreatic Cystic Lesions in Confocal Laser Endomicroscopy Using Patch-Based Image Segmentation

The early identification of pancreatic cystic lesions plays a vital part in the treatment of patients diagnosed with pancreatic cancer. However, it continues to provide a significant difficulty. This study employs the VGG19 network to construct a deep-learning model aimed at predicting the specific type of pancreatic cyst. The dataset utilized for training consists of 127,332 picture patches derived from five distinct types of pancreatic cystic videos. The training images are preprocessed using Gaussian filtering and an image patch segmentation scheme. Data augmentation is achieved by rotating the circular component in the training images. During the testing phase, a Gaussian filtering approach is applied to the test video as a preprocessing step prior to classification. The image patch segmentation scheme is also employed throughout the testing phase of our study. Our proposed methodology has the capability to autonomously categorize the specific feature type of pancreatic cystic in the test videos, while simultaneously documenting the prediction outcomes on a frame-by-frame basis. The methodology was assessed using 18 test videos, including a total of 11,059 frames. The experimental results demonstrate that the proposed methodology achieves a classification accuracy of up to 83% for different types of pancreatic cysts.