How to load a single image in MONAI for segmenation?

[Adwaitt]

Hello everyone, hope you are having a good day :)

I wanted to ask how can I load just a single image for segmentation in MONAI. I trained a SegResNet on a nifti image of shape (h, w, number of images = 155) and it stacks the image (4 images) therefore the final output is (4, h, w, number of images = 155) I just want to load only a single image (4, h, w, 1) how can I do that?

Also if I just want to load one single image (no nifti say a png of shape (224, 224)) to meet the requirement of channels is it okay if I just stack the image with just zeros i.e np.dstack((img, np.zeros(h, w), np.zeros(h, w), np.zeros(h, w)))?

Thank you very much for you time :D

[Nic-Ma]

Hi @Adwaitt ,

I think the reason it stacked the images is that you passed a list of file paths to LoadImage transform, if you only pass 1 path, it will only load 1 image.

Thanks.

[Adwaitt]

Thanks for your response Nic. What I meant was, I'm not loading the path via any transform function. The model initially took a nifti volume of size (batch_size, number_of_stacked_images (channels), height, width, number_of_images). Here what I wanna do is, (batch_size, number_of_stacked_images, height width, 1). While I loaded it and fed it to my model it did run but instead of getting 1 output for the segmentation model I got 8 and the mask was not clear at all. (1, 4, 224, 224, 8). Idk why this happened for a single case when I tried it for 144 images the output was (1, 4, 224, 224, 144) and the mask was clear too.

[Nic-Ma]

Sorry I don't quite understand your question. Maybe I missed some points, could you please help show your question with some example?
@rijobro @ericspod @wyli Can you guys help share some comments?

Thanks in advance.

[Adwaitt]

img = nib.load('path to nib image')
r = img[:, :, 100] #Loading the 100th slice from a nifti image of 240 slices
r = tfms(image = r)['image'] #Applying my transformation
stacker = np.zeros((224, 224))
r = {'image': r, 'label': None}#dictionary to apply MONAI transforms
r = new_transform(r)['image']#New_transform = MONAI transform
r = np.dstack((r, np.zeros((224, 224)), np.zeros((224, 224)), np.zeros((224, 224))))# Stacking zeros along depth.`

r = r[np.newaxis, ...]# Running this code twice to add two new axis so as to make the input compatible with the model
r = r.transpose(0, -1, 2, 3, 1) #Transposing the image so as to make it the desired shape

with torch.no_grad():
n = model.forward(torch.tensor(r, dtype = torch.float).cuda())
n = np.squeeze(n.cpu().numpy(), 0)
n.shape

Output for a single slice of nifti image (240, 155, 240): (this is what I desire)

Output for single slice of image(240, 155, 1): (this is what I get)

Also the output for this single image contains 8 images. Output is (224, 224, 8) (I don't understand why is that)

[Adwaitt]

Eric, I'm unable to "Mark as an answer" button to your reply.

[Adwaitt]

Also Thank you Nick and Eric :D MONAI is the best

[Nic-Ma]

I marked the thread as the answer.

Thanks.

[Adwaitt]

Oh, Thank you Nic and Thank you Team MONAI.

[Nic-Ma]

You are welcome.