An automatic cascaded approach for pancreas segmentation via an unsupervised localization using 3D CT volumes

Automatic organ segmentation using computed tomography (CT) images can support radiologists while carrying out quantitative and qualitative analyses of various types of cancer in their early stages. This work is aimed at automating the segmentation of the pancreas using CT images, which would ultimately aid in the early detection of pancreatic cancer. The pancreas is a small and challenging organ for automatic segmentation due to its variability in shape, size, and position. The state-of-the-art convolution neural networks (CNNs) based approaches have reported acceptable outcomes for stable large organs, but limited results for small organs like the pancreas. Although CNNs based results are promising, they utilized the supervised approach for localization, which required annotations. Hence, to avoid the need for annotations during localization, a novel unsupervised localization approach is proposed. The proposed approach localizes the pancreas from 3D CT volume using the spatial locations of stable large organs such as the liver and spleen. However, their spatial locations are detected in an unsupervised way. Furthermore, a 2D multi-view fusion deep learning model is used to extract the boundaries of the pancreas using the small bounding box around the pancreas region. The segmentation results are very encouraging and motivating to use an unsupervised localization approach instead of a supervised approach. A large number of experiments are performed using the NIH-82 CT dataset, which reveals that the proposed localization approach can achieve good segmentation results.