Rui Liu

you can use ITK-Snap to load the segmentation , then it can visualize it.

> > In line 74 (gt_im = np.swapaxes(self.train_im, 0, 1)[4]) at extract_patches.py I will get the following error. > > ValueError: bad axis2 argument to swapaxes. > > This is...

My point : for the Brats Dataset, the input does have 4 channel, ie. t1, t1ce, t2, flair, so I think it's correct, no bug.

> For Brats Dataset, it is OK. But, for another dataset, the channel may be not 4. So, `self.Unet = Unet(in_channels=num_channels, base_channels=16, num_classes=4)` may be more robust. yes， agree with...

> > > > 你可以修改这行代码 > > > > prepare3dtraindata(pathlgg_list, bratslgg_path, trainImage, trainMask, (64, 128, 128), 3, 15, 2) > > > > 把他修改为 > > > > prepare3dtraindata(pathlgg_list, bratslgg_path,...

相当于“去掉最高分和去掉最低分”:)

> Yes, it's ok. This operator makes x_start in [-1, 1] range, bacause the diffusion model's input is usually in [-1, 1]. Thanks for your quick reply. Another question: I...

> No. You can train the model directly, and this operator only has a little change on the validation results. Thanks！