3D CoordConv Segmentation

Grand Challenge 2017 Multi-Modality Whole Heart Segmentation

• http://www.sdspeople.fudan.edu.cn/zhuangxiahai/0/mmwhs/

Contribution

• An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution [https://arxiv.org/abs/1807.03247]

Training Run

In code directory

> python main.py --params=ct_train.json

Result

Model	Background	MLV	LABC	LVBC	RABC	RVBC	ASA	PUA	Average DSC
U-net 3D	0.995	0.918	0.929	0.912	0.925	0.923	0.843	0.923	0.909
U-net 3D + CoordConv	0.995	0.919	0.926	0.912	0.933	0.924	0.928	0.897	0.920

• MLV: the Myocardium of the left ventricle, LABC: the left atrium blood cavity, LVBC: the left ventricle blood cavity, RABC: the right atrium blood cavity, RVBC: the right ventricle blood cavity, ASA: the ascending aorta, PUA: the pulmonary artery
• Average DSC is average of classes that excluded background

Left: Mask, Middle: U-Net 3D, Right: U-Net 3D CoordConv

Details

Data	Number of train set	Number of validation set	Patch dim	Resize rate	Batch size	Epochs	Number of train patch image	Number of validation patch image	Metric	Loss function	Optimizer	Learning rate	Number of GPU
CT	18	2	96	0.7	2	100	20	100	Dice Similarity Coefficient	dice coefficient loss	Adam	0.0001	4

Limit

The host server is down, so the test set can no longer be evaluated.