The low resolution of thick CT/MR with large spacing will lead to misdiagnosis and bring a huge difficulty to the automatic organ or lesion segmentation. Especially in automatic pancreatic tumor segmentation, due to the limitation of equipment, the tumors only exist in a few slices, and the continuity between slices of 3D CT/MR image is poor. Besides, the features of tumors, such as the size, shape, location, intensity and so on, vary dramatically according to different cases. The fuzzy boundaries of small tumors also bring great uncertainty to the segmentation task. Aiming to solve those problems, we introduce the LLE-based interpolation neural network into the pancreatic tumor segmentation task, which mainly includes the following improvements: 1) We utilize local linear embedding (LLE) to model the relationship between adjacent slices and the interpolated slice. It adapts the spatial transformation of the organ between slices. 2) Neural network, combining with the LLE module, is designed to significantly enhance the image resolution, thus can generate more continuous and clearer images for each sequence. 3) Multiscale cascade strategy is adopted in the network to reduce the influence of drastic changes in tumor size on segmentation results. Experiments are carried out in MR images with 3.5mm thickness provided by Changhai hospital, and CT images on Medical Segmentation Decathlon pancreatic tumor challenge, respectively. The results show that our proposed model has a sensitivity of 96.55% and a dice coefficient of 62.8% (±25.8%) on the MR dataset, and a dice coefficient of 50.6% (±,30.9%) on the CT dataset.

Donec vel ipsum vulputate, dapibus magna quis, auctor mauris. Aliquam ut dapibus risus. Duis sit amet magna posuere, interdum metus sed, maximus tortor. Maecenas tempus ut magna nec facilisis. Pellentesque vel ultrices nunc, eget ultrices mauris. Ut eget ligula quis purus porta rhoncus eget at eros. In id felis ullamcorper, vehicula ante sed, cursus magna. Curabitur libero nibh, posuere in urna in, viverra pulvinar lectus. Pellentesque rutrum tempus elit at hendrerit. Sed enim est, bibendum in augue iaculis, viverra venenatis sapien. Nullam sagittis augue maximus mi interdum maximus. Praesent a cursus massa.

Duis vulputate odio vitae consectetur lacinia. Vivamus tellus nisi, hendrerit eget tortor eget, rutrum lobortis tellus. Donec convallis scelerisque egestas. Vestibulum ex urna, commodo at interdum vitae, malesuada id arcu. Praesent et enim libero. Proin et felis nisi. Duis imperdiet placerat dapibus.