(Tuesday, 7/16/2019) 11:00 AM - 12:15 PM
Room: Stars at Night Ballroom 2-3
Purpose: To take the advantages of less computation and memory space from 2D network, and the 3D information of neighbor slices from 3D network, we propose a novel mixing 2D-3D Fully Convolutional Neural Network (FCN) for pancreas auto-segmentation.
Methods: A fully-annotated pancreas segmentation was built based on FCN and batch processing technique. The network structure consisted of 2D and 3D components, the typical shape of 2D feature map was reshaped to [1, 16, weight, height, channel], so that 3D convolution can be performed and connections between neighboring slices were considered.Image batches were sampled to make sure that neighbor slices in the batch are also neighbors in the original 3D volume. We randomly selected one slice, then 16 slices centered on the selected slice were sampled as an image batch, which made the 2D image batch in fact a 3D volume data with batch size as 1. Data augmentation was implemented on the whole image batch simultaneously but not each slice in the batch respectively. Those image batches were fed to the new network and the 2D-3D Mixture FCN to train end to end. The algorithm was trained and tested on 82 abdomen cases whose pancreas were manually delineated. Of these, 50 cases were used for training and 32 cases were used for testing. The segmentation accuracy was evaluated by Dice Similarity Index (DSI).
Results: The segmentation results yielded by the proposed method was compared to that of the 2D FCN. The average DSIs were 0.715 and 0.706 for the 2D-3D mixture method and the 2D FCN, respectively.
Conclusion: The proposed method takes the 3D information of CT volume into account with the original high resolution undiminished. The results demonstrate that the 2D-3D-mixture FCN is effective and achieved better segmentation performance over the 2D FCN.
Funding Support, Disclosures, and Conflict of Interest: S. Tan and B. Ye are supported in part by the National Natural Science Foundation of China, under Grant Nos. 61375018 and 61672253.
Keywords
Not Applicable / None Entered.
Taxonomy
Contact Email
