Separating outputs for multi-label segmentation

[ghost]

Hi,
I have labels for two features I am segmenting and for later processing it would be useful to export the model segmentations separately. If I'm missing a tutorial/example for how to do this please let me know but here is what I've tried so far:

The shape of the whole segmentation is:

torch.Size([1, 2, 64, 64, 64])

Therefore I have tried iterating over the second element to produce separated outputs with:

model.load_state_dict(torch.load(trained_model))
model.eval()
    with torch.no_grad():
        for image in range(len(case_name)):
            print(f'Processing {case_name[image]}')

            val_input = val_ds[image]["image"].unsqueeze(0).to(device)
            post_trans = Compose([Activations(sigmoid=True), AsDiscrete(threshold_values=True, logit_thresh=0.7)]     )  
    
            val_output = model(val_input)
            val_output_bin = post_trans(val_output)

             # separate outputs for the two labels for features LP and TY
             feature = ['LP', 'TY']
             for anatomy in range(2):
                  print(val_output_bin_largest.shape)
                  saver = NiftiSaver(output_dir=outputdata_dir+ f'/{case_name[image]}', output_postfix=f'{feature[anatomy]}_seg', dtype=(['numpy.float16']))
                  saver.save_batch(val_output_bin[:, anatomy, :, :, :], meta_data = {'filename_or_obj': case_name})

This does output the labels separately, but in a different orientation to the un-separated output saver.save_batch(val_output_bin) and this orientation is not ideal for future processing steps. I know I could do some transforms but am I missing something? I had hoped the separate segmentations would simply output in the same orientation as separate files.

Edit:
I realised it would be useful to share the different output shapes.
Combined output:(64, 64, 64, 2)
Separated outputs: (64, 64, 1, 64)

Any assistance would be greatly appreciated as always! Thank you in advance

[Nic-Ma]

Hi @jnc13 ,

Thanks for your interest here.
I didn't quite understand your question, but I think maybe you want to add argmax to the segmentation, you can set it by AsDiscrete transform.
If you want to split channels, you can use SplitChannel transform.
@rijobro please feel free to correct me if I misunderstand anything here.

Thanks.

[ghost]

Yep, that solved it. Now working exactly as I hoped it would.

Thank you so so much @Nic-Ma , not just for answering this question but your continual help and support. It makes MONAI so much more approachable for those of us who want to apply the technology to our medical research but don't have a background in coding. So thank you :)

[Nic-Ma]

No problem, glad to see your good progress.

Thanks.

[oscarramirezs]

hi Nic
Im trying to fo the same

  vali_transforms = Compose(
      [
          LoadImaged(keys=["image"]),
          EnsureChannelFirstd(keys=["image"]),
          Orientationd(keys=["image"], axcodes="RAS"),      
          EnsureTyped(keys=["image"]),
      ]
  )


with torch.no_grad():
        for val_data in vali_loader:
            
            val_inputs = val_data["image"].to(device)
            
            roi_size = (512, 512, 256)
            
            sw_batch_size = 1
            
            val_outputs = sliding_window_inference(
                val_inputs, roi_size, sw_batch_size, model_scapula_humerus)
            
            val_outputs = [post_trans(i) for i in decollate_batch(val_outputs)]
            
            val_output_bin_largest_split = SplitChannel(channel_dim=1)(val_outputs)
            
            for i, c in enumerate(val_output_bin_largest_split):
                saver = NiftiSaver(output_dir="./segmentation_output", output_postfix=f'class_{i}')
                saver.save_batch(c)

but I got the following erreur

  AttributeError                            Traceback (most recent call last)
  <ipython-input-76-acde0dae54fd> in <module>
        7                 val_inputs, roi_size, sw_batch_size, model_scapula_humerus)
        8             val_outputs = [post_trans(i) for i in decollate_batch(val_outputs)]
  ----> 9             val_output_bin_largest_split = SplitChannel(channel_dim=1)(val_outputs)
       10             for i, c in enumerate(val_output_bin_largest_split):
       11                 saver = NiftiSaver(output_dir="./segmentation_output", output_postfix=f'class_{i}')
  
  C:\ProgramData\Anaconda3\lib\site-packages\monai\transforms\utility\array.py in __call__(self, img)
      295 
      296     def __call__(self, img: NdarrayOrTensor) -> List[NdarrayOrTensor]:
  --> 297         num_classes = img.shape[self.channel_dim]
      298         if num_classes <= 1:
      299             raise RuntimeError("input image does not contain multiple channels.")
  
  AttributeError: 'list' object has no attribute 'shape'

thank you for your help

Oscar

[KumoLiu]

Hi @oscarramirezs, decollate_batch is to de-collate a batch of data, so put it out of loop like this:

for val_data in vali_loader:
            
    val_inputs = val_data["image"].to(device)
            
    roi_size = (512, 512, 256)
            
    sw_batch_size = 1
            
    val_outputs = sliding_window_inference(
        val_inputs, roi_size, sw_batch_size, model_scapula_humerus)
            
val_outputs = [post_trans(i) for i in decollate_batch(val_outputs)]
            
val_output_bin_largest_split = [SplitChannel(channel_dim=1)(i) for i in val_outputs]

[oscarramirezs]

Hi @KumoLiu thank you for your help

with torch.no_grad():
        for val_data in vali_loader:
            
            val_inputs = val_data["image"].to(device)
            
            roi_size = (512, 512, 224)
            
            sw_batch_size = 1
            
            val_outputs = sliding_window_inference(
                val_inputs, roi_size, sw_batch_size, model_scapula_humerus)
            
        val_outputs = [post_trans(i) for i in decollate_batch(val_outputs)]
        
            
        val_output_bin_largest_split = [SplitChannel(channel_dim=1)(i) for i in val_outputs]
 
        for i, c in enumerate(val_output_bin_largest_split):
            saver = NiftiSaver(output_dir="./segmentation_output", output_postfix=f'class_{i}')
            saver.save_batch(c)

I don't get a batch but a file for each image.
you should get only two files. and thank you again

file written: segmentation_output\0\0_class_0.nii.gz.
file written: segmentation_output\1\1_class_0.nii.gz.
file written: segmentation_output\2\2_class_0.nii.gz.
file written: segmentation_output\3\3_class_0.nii.gz.
file written: segmentation_output\4\4_class_0.nii.gz.
file written: segmentation_output\5\5_class_0.nii.gz.
file written: segmentation_output\6\6_class_0.nii.gz.
file written: segmentation_output\7\7_class_0.nii.gz.
file written: segmentation_output\8\8_class_0.nii.gz.
file written: segmentation_output\9\9_class_0.nii.gz.
file written: segmentation_output\10\10_class_0.nii.gz.
file written: segmentation_output\11\11_class_0.nii.gz.
file written: segmentation_output\12\12_class_0.nii.gz.
file written: segmentation_output\13\13_class_0.nii.gz.
file written: segmentation_output\14\14_class_0.nii.gz.
file written: segmentation_output\15\15_class_0.nii.gz.
file written: segmentation_output\16\16_class_0.nii.gz.
file written: segmentation_output\17\17_class_0.nii.gz.
file written: segmentation_output\18\18_class_0.nii.gz.
file written: segmentation_output\19\19_class_0.nii.gz.
file written: segmentation_output\20\20_class_0.nii.gz.
file written: segmentation_output\21\21_class_0.nii.gz.
file written: segmentation_output\22\22_class_0.nii.gz.
file written: segmentation_output\23\23_class_0.nii.gz.
file written: segmentation_output\24\24_class_0.nii.gz.
file written: segmentation_output\25\25_class_0.nii.gz.
file written: segmentation_output\26\26_class_0.nii.gz.
file written: segmentation_output\27\27_class_0.nii.gz.
file written: segmentation_output\28\28_class_0.nii.gz.
file written: segmentation_output\29\29_class_0.nii.gz.
file written: segmentation_output\30\30_class_0.nii.gz.
file written: segmentation_output\31\31_class_0.nii.gz.