question about co-training models

[guopengf]

In your paper. you mentioned the networks co-trained in 3DSeg-8 dataset, including Brats. Would like lease more detail about co-training? In other words, Brats dataset contains T1, T2, T1c, Flair (4 different weighted sequences) for each patient. And the annotation is also based on all 4 sequences. Therefore, the input is supposed to be 4 channels. However, the input is usually only one channel for CT dataset. So, how did you deal with this situation? re-initializing first conv layer?