3D UNETR model

[abhishekpatil32]

I am very new to MONAI and using the 3D UNETR model to train the model and to segment a small region of interest in a CT scan.

What changes are required in the code to run the unetr_btcv_segmentation_3d.ipynb code ??
link: [https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/unetr_btcv_segmentation_3d.ipynb]

I did not find any tutorials for binary segmentation.

I really appreciate any help you can provide.

[KumoLiu]

Hi @abhishekpatil32, thanks for your interest here.

You can refer to the spleen segmentation here which is about binary segmentation: https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/spleen_segmentation_3d.ipynb
You need adjust the output channel, activation function and loss accordingly.

Hope it helps,
Thanks.

[abhishekpatil32]

Thank you for your answer, @KumoLiu . I worked around the issue and successfully ran the UNETR model with the parameters mentioned above. However, the model's performance is poor, with a very low mean Dice score (~0.05).

Could the issue be related to the input data? My target area for segmentation is very small (a 15 mm cube), and the input data consists of CT images and labels (both in .nii.gz format). I am wondering if preprocessing the labels is necessary before training.

Please suggest measures to improve the Dice score.
Thank you in advance

[abhishekpatil32]

Hello @KumoLiu, when I try to run the script for visualization named /unetr_btcv_segmentation_3d_lightning.ipynb

slice_map = {
    "img0035.nii.gz": 170,
    "img0036.nii.gz": 230,
    "img0037.nii.gz": 204,
    "img0038.nii.gz": 204,
    "img0039.nii.gz": 204,
    "img0040.nii.gz": 180,
}
case_num = 4
net.load_from_checkpoint(os.path.join(root_dir, "best_metric_model.ckpt"))
net.eval()
net.to(device)

with torch.no_grad():
    img_name = os.path.split(net.val_ds[case_num]["image"].meta["filename_or_obj"])[1]
    img = net.val_ds[case_num]["image"]
    label = net.val_ds[case_num]["label"]
    val_inputs = torch.unsqueeze(img, 1).cuda()
    val_labels = torch.unsqueeze(label, 1).cuda()
    val_outputs = sliding_window_inference(val_inputs, (96, 96, 96), 4, net, overlap=0.8)
    plt.figure("check", (18, 6))
    plt.subplot(1, 3, 1)
    plt.title("image")
    plt.imshow(val_inputs.cpu().numpy()[0, 0, :, :, slice_map[img_name]], cmap="gray")
    plt.subplot(1, 3, 2)
    plt.title("label")
    plt.imshow(val_labels.cpu().numpy()[0, 0, :, :, slice_map[img_name]])
    plt.subplot(1, 3, 3)
    plt.title("output")
    plt.imshow(torch.argmax(val_outputs, dim=1).detach().cpu()[0, :, :, slice_map[img_name]])
    plt.show()

I get a strange error
AttributeError: 'Net' object has no attribute 'val_ds'

Has anyone resolved this ?

[KumoLiu]

The error AttributeError: 'Net' object has no attribute 'val_ds' indicates that Python is trying to access an attribute named val_ds on an object of type Net, but the Net class (or its instances) does not have an attribute with that name. So if you didn't define val_ds in your net, please check your code.