Question about how to choose the crop area and train the model

[Tristacheng]

Hi, I want to implement your code to train on my data but meet some hinders, I'd appreciate if you release the code about how to train you model. Also, I want to know how do you choose the precrop area, or you just choose area larger than the hippocampus area?

[bthyreau]

Hi, thanks ! You can look at the two MNI152_*.nii.gz to appreciate the FOV that is used in the program; That box it is further splitted in half (with Right sides mirrored) before finally entering into the model. If you want to retrain your own model, i suppose any bouding-box area which encompass the whole hippocampus would work. The main constraint here is the memory resources (3d volumes grow large quickly), although it of course depends on the resolution chosen. I recommend using a box large-enough to accomodate for possible registration inaccuracies at inference time. For further robustness, you could even introduce purposedly poorly-registered training examples. 2017-10-31 22:50 GMT+09:00 Tristacheng <notifications@github.com>:

…

Hi, I want to implement your code to train on my data but meet some hinders, I'd appreciate if you release the code about how to train you model. Also, I want to know how do you choose the precrop area, or you just choose area larger than the hippocampus area? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAkESaeBvUln0sHj7ZkjAlcYQM-aYG7Hks5sxyWigaJpZM4QMxQC> .