save predict volume

[mikiwang820]

Hi,

I wrote an evaluation to predict my labels for segmentation task and save as nifti file, but some outputs look like being shift.
Additionally, I got around 0.8 dice score for one patient before doing "saver.save_batch" so that I think the predicational volume's position is correct at this moment.
After this, I saved it and calculated dice score with ground truth label and prediction, and now dice score turns to 0.57 because of image shift.
Is there any problem when I applied "saver.save_batch?
I also attach my evaluation code and how I calculate dice score.
I will appreciate your any help.

1. Evaluation code

crop_size = (96, 96, 32)
val_transforms = Compose(
    [
        LoadImaged(keys=["img", "seg"]),
        AddChanneld(keys=["img", "seg"]),
        Orientationd(keys=["img", "seg"], axcodes="RAS"),
        ScaleIntensityRanged(
            keys=["img"], a_min=-200, a_max=250, b_min=-1.0, b_max=1.0, clip=True,
        ),
        CropForegroundd(keys=["img", "seg"], source_key="img"),
        ToTensord(keys=["img", "seg"]),
    ]
)

val_ds = CacheDataset(data=val_files, transform=val_transforms, cache_rate=1.0, num_workers=4) 
val_loader = DataLoader(val_ds, batch_size=1, num_workers=4) 

dice_metric = DiceMetric(include_background=True, reduction="mean")
post_trans = Compose([Activations(sigmoid=True), AsDiscrete(threshold_values=True)])

device = torch.device("cuda:0")
model = UNet(
    dimensions=3,
    in_channels=1,
    out_channels=2,
    channels=(16, 32, 64, 128, 256),
    strides=(2, 2, 2, 2),
    num_res_units=2,
    norm=Norm.BATCH,
).to(device)

post_pred = Compose([AsDiscrete(argmax=True), KeepLargestConnectedComponent(applied_labels=[1], connectivity=1)])
output_dir = os.path.join(data_dir, "output")
model.load_state_dict(torch.load("/lits_1mm/monai_models/for_evaluation.pth"))
model.eval()

with torch.no_grad():
    metric_sum = 0.0
    metric_count = 0
    saver = NiftiSaver(output_dir)
    for val_data in val_loader:
        val_images, val_labels = val_data["img"].to(device), val_data["seg"].to(device)
        roi_size = crop_size
        sw_batch_size = 1
        #Evaluate
        val_outputs = sliding_window_inference(
            val_data["img"].to(device), roi_size, sw_batch_size, model
        )
        val_outputs = post_pred(val_outputs)
        #val_labels = post_label(val_labels)
        value, _ = dice_metric(y_pred=val_outputs, y=val_labels)
        metric_count += len(value)
        metric_sum += value.item() * len(value)
        saver.save_batch(val_outputs, val_data["img_meta_dict"])
    metric = metric_sum / metric_count
    print("evaluation metric:", metric)

2. My dice score code

import os
import glob
import SimpleITK as sitk
import numpy as np

dice_sum = 0

def binary_dice3d(s,g):
    #dice score of two 3D volumes
    num=np.sum(np.multiply(s, g))
    denom=s.sum() + g.sum() 
    if denom==0:
        return 1
    else:
        return  2.0*num/denom
    
GT_route = "/Users/wangmeiqi/code_and_data/HCC/temp/true"
predict_route = "/Users/wangmeiqi/code_and_data/HCC/temp/predict"
GT_images = sorted(glob.glob(os.path.join(GT_route, "*.nii.gz")))
predict_images = sorted(glob.glob(os.path.join(predict_route, "*.nii.gz")))
save_route = "/Users/wangmeiqi/code_and_data/HCC/temp"


for i in range(len(GT_images)):
    
    read_gt = sitk.ReadImage(GT_images[i])
    GT_array = sitk.GetArrayFromImage(read_gt)
    predict_array = sitk.GetArrayFromImage(sitk.ReadImage(predict_images[i]))
    dice = binary_dice3d(predict_array, GT_array)
    print(predict_images[i].split("/")[-1] + ": " + str(dice))
    dice_sum = dice_sum + dice

avg_dice = dice_sum/len(GT_images)

3. Output volume

[rijobro]

Am I right in thinking that you crop your segmentations to a size of 96x96x32, yet (I'm guessing here) you superpose that on top of your original, non-cropped image? If so, this will be where the difference is coming in.

[mikiwang820]

Sorry, I think I misunderstood before. However, I check my images and labels again before doing post transfer. As shown in the below figure, the image and label are matched. Therefore, I guess the problem maybe comes from the post transform or I didn't use post transform functions in a right way?

1. the brighter one is predicted label for different patients
   

[rijobro]

Hi, yes in that case it looks like the discrepancy is due to writing the segmentations at the end. Possibly related to this issue in that case: #365.

Earlier you sent a few screenshots from ITKSnap, but it's missing pixdim[1] and pixdim[2] (i.e., voxel spacing in x- and y-directions), and also the number of voxels in all directions. Would you be able to send that information?

[mikiwang820]

Thank you so much @rijobro
Here is the information for pixdim[1] and pixdim[2].

1. image

2. ground truth label

3. predicted label

[mikiwang820]

Hi @rijobro ,
I finally find my problem.
I use "monai.data.Dataset" instead of "CacheDataset", it works.
Thanks for your help.

[rijobro]

I'm surprised that solves the problem, but glad that it does!