Improving 3D segmentation mean dice

[eanemo]

Hi,

I am trying to train a model for a binary 3D segmentation task. I have created a custom dataset that includes 46 volumes (nii.gz format) for training and 9 for testing. I have used the example provided in the tutorial section as the base code for training.
I have used several networks architectures (UNet, VNet, SegResNet, etc), solvers (Adam, AdamW, etc) and values for learning rate param (0.001, 0.0001, etc). The best model that I have obtained so far has a mean dice value of 0.65 and loss is about 0.12.
I have tested the source code by using a testing dataset created using the create_test_image_3d method from Monai. Using this testing dataset I get a segmentation model with a mean dice value over 0.95 without any problems. So I presume that the source code is correct.
My custom dataset contains images like the following ones:

Image

Segmentation (Ground Truth)

The model obtain a segmentation like this one:

Any idea, suggestion for improving my current results?

[Nic-Ma]

Hi @dongyang0122 @wyli ,

Could you please help share some advice about this ticket?
I feel the image & label resolution is too low and cause the model hard to predict.

Thanks.

[eanemo]

The image resolution used in this training process is 256x256x256. In the examples directory, I have seen some code that uses smaller image resolution (128x128x128) without any problem. So, I think that resolution of the image in the dataset should not be the problem.

[dongyang0122]

The "resolution" in the context of medical imaging denotes spacing (physical size of pixels/voxels). You can want to check the transform SpacingD to further increase it.

[eanemo]

Ok. I will try to increase the resolution of the image to check if this is the problem.

[dongyang0122]

Hi @eanemo, can you share the usage of create_test_image_3d or parts of implementation?

[eanemo]

Hi, the code that generates the testing dataset is as following:

for i in range(20):
    img_filename = "image_%d_cut.nii.gz" % (i)
    label_filename = "label_%d_segmentation.nii.gz" % (i)
    noisy_im, label = create_test_image_3d(height=128, width=128, depth=128, noise_max=0.1, num_seg_classes=1)
    nii_im = nib.Nifti1Image(noisy_im, np.eye(4))
    nii_label = nib.Nifti1Image(label, np.eye(4))
    nib.save(nii_im, join(target_dir, img_filename))
    nib.save(nii_label, join(target_dir, label_filename))

[dongyang0122]

What do generated image and label look like?

[eanemo]

Images generated with the code above looks like the following ones:

Image

Segmentation