problem about 3dhighresnet

[shieldlance]

I'm using niftynet to do some work about heart CT segmentation. And i used highres3d_net. The training was perfect. However when I applied the trained network to some new images，the output was terrible. one image used for training is illustrated in 1, and its output in2. the new image is in3,and the output 4.
How should I do to get an accurate result?
Any help would be appreciated.
image for training
output of above
new image
output of new image

[Zach-ER]

Are the images from the same machines? I.e. is the test data acquired in the same way as the training data?