abdominal multi-organ segmentation using pytorch,pytorch version: 0.4.0

the data come from an online challenge called Multi-atlas labeling Beyond the Cranial Vault, for the detail, you can check this link:https://www.synapse.org/#!Synapse:syn3193805/wiki/217752. in this challenge, the task is to segement 13 different kind of organ as follow:

i use the trainging set given by the competition organizer. The training set include 30 CT data.I randomly divided it into 25 for training and 5 for evaluation. and organize them as follow:

i normalized the axial spacing to 3mm. and truncated the hu value to a certain range. only the slice contain organ are used to train the network.

i use two u-shape like 3D FCN, and add residual connection at a certain group of convlayers. In order to increase the receptive field，i add some hybrid dilated convlayer to the last two stage of the encoder.most idea come form [1].

i use adam optim and set the initial learning rate to 1e-4, train on three GTX 1080TI with batch size equal to three.the whole trainging process take about 13 hours.

i use mean dice coefficient as metrics.

i have implement different kind of loss function, you can try which one work best in your data.

Here is the best of the above results:

you can copy the value in bset_result.xlsx to show.xlsx to get the above picture

• other loss function
• data augmentation

1. Roth H R, Shen C, Oda H, et al. A multi-scale pyramid of 3D fully convolutional networks for abdominal multi-organ segmentation[J]. arXiv preprint arXiv:1806.02237, 2018.

2. Milletari F, Navab N, Ahmadi S A. V-net: Fully convolutional neural networks for volumetric medical image segmentation[C]//3D Vision (3DV), 2016 Fourth International Conference on. IEEE, 2016: 565-571.

3. Fidon L, Li W, Garcia-Peraza-Herrera L C, et al. Generalised wasserstein dice score for imbalanced multi-class segmentation using holistic convolutional networks[C]//International MICCAI Brainlesion Workshop. Springer, Cham, 2017: 64-76.

4. Sudre C H, Li W, Vercauteren T, et al. Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations[M]//Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Springer, Cham, 2017: 240-248.