Questions about inference results

[Alex8211101]

have tried two methods to generate the predicted results. But there are some problems.
n this section of the program there is a problem that the image does not match the origin of the result.
`test_dicts = [{'image': image_name, 'label': label_name}
for image_name, label_name in zip(test_images, test_segs)]

# define transforms for image and segmentation
val_transforms = Compose(
    [
        LoadImaged(keys=["image", "label"]),
        EnsureChannelFirstd(keys=['image','label']),
        ScaleIntensityd(keys="image"),
        EnsureTyped(keys=["image", "label"]),
    ]
)
val_ds = monai.data.Dataset(data=test_dicts, transform=val_transforms)
# sliding window inference need to input 1 image in every iteration
val_loader = DataLoader(val_ds, batch_size=1, num_workers=0, collate_fn=list_data_collate)
dice_metric = DiceMetric(include_background=False, reduction="mean", get_not_nans=False)
post_trans = Compose([EnsureType(), Activations(softmax=True), AsDiscrete(argmax=True),])
# post_trans = Compose([EnsureType(), Activations(softmax=True), AsDiscrete(threshold_values=True, logit_thresh=0.5)])
saver = SaveImage(output_dir="./out/SegResNet/tibia", output_ext=".nii", output_postfix="seg")
# try to use all the available GPUs
devices = [torch.device("cuda" if torch.cuda.is_available() else "cpu")]
model = monai.networks.nets.SegResNet(out_channels=2).cuda()

model.load_state_dict(torch.load(os.path.join('log/SegResNet/tibia', "best_metric_model.pth")))

# if we have multiple GPUs, set data parallel to execute sliding window inference
if len(devices) > 1:
    model = torch.nn.DataParallel(model, device_ids=devices)

model.eval()
with torch.no_grad():
    for val_data in val_loader:
        val_data['image'], val_labels = val_data["image"].to(devices[0]), val_data["label"].to(devices[0])
        # define sliding window size and batch size for windows inference
        # val_images = torch.unsqueeze(val_images,dim=0)
        roi_size = (320, 320,128)
        sw_batch_size = 4
        val_outputs = sliding_window_inference(val_data['image'], roi_size, sw_batch_size, model,overlap=0.7)
        val_outputs = [post_trans(i) for i in decollate_batch(val_outputs)]
        val_labels = decollate_batch(val_labels)
        meta_data = decollate_batch(val_data["image_meta_dict"])
        # compute metric for current iteration
        dice_metric(y_pred=val_outputs, y=val_labels)
        for val_output, data in zip(val_outputs, meta_data):
            saver(val_output, data)
    # aggregate the final mean dice result
    print("evaluation metric:", dice_metric.aggregate().item())
    # reset the status
    dice_metric.reset()`

[Alex8211101]

In this part of the program, there is a problem that the result is displayed incorrectly.

test_images = sorted(glob(os.path.join(opt.images_folder, 'test', 'image*.nii')))
test_dicts = [{'image': image_name}for image_name in test_images]
test_org_transforms = Compose(
[
LoadImaged(keys="image"),
EnsureChannelFirstd(keys='image'),
Orientationd(keys=["image"], axcodes="RAS"),
EnsureTyped(keys="image"),
]
)
test_org_ds = Dataset(
data=test_dicts, transform=test_org_transforms)
test_org_loader = DataLoader(test_org_ds, batch_size=1, num_workers=0)
post_transforms = Compose([
EnsureTyped(keys="pred"),
Activationsd(keys="pred",softmax=True),
Invertd(
keys="pred",
transform=test_org_transforms,
orig_keys="image",
meta_keys="pred_meta_dict",
orig_meta_keys="image_meta_dict",
meta_key_postfix="meta_dict",
nearest_interp=False,
to_tensor=True,
),
AsDiscreted(keys="pred",to_onehot=2,argmax=True,threshold=0.5),
SplitChanneld(keys="pred", channel_dim=0),
SaveImaged(keys="pred_0",
meta_keys="pred_meta_dict",
output_dir="./out/SegResNet/tibia/",
output_postfix="seg_0",output_ext='.nii' ,resample=False),
SaveImaged(keys="pred_1",
meta_keys="pred_meta_dict",
output_dir="./out/SegResNet/tibia/",
output_postfix="label", output_ext='.nii.gz', resample=False),
])
net = monai.networks.nets.SegResNet(out_channels=2)

net.cuda()
net.load_state_dict(torch.load(os.path.join('log/SegResNet/tibia/', "best_metric_model.pth")))
net.eval()
with torch.no_grad():
for test_data in test_org_loader:
test_data['image'] = test_data["image"].cuda()
roi_size = (256,256,96)
sw_batch_size = 4
test_data["pred"] = sliding_window_inference(
test_data['image'], roi_size, sw_batch_size, net)
test_data = [post_transforms(i) for i in decollate_batch(test_data)]

[Nic-Ma]

Hi @Alex8211101 ,

I think maybe you need to add the invertd transform for the postprocessing, referring to:
https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/torch/unet_inference_dict.py#L70

Thanks.

[Alex8211101]

Thanks for your quick reply,I have used Invert,when I tried to use Orientationd(keys="image", axcodes="RAS"),in pre_transforms,I got the result like the picture below,is it related to Orientationd?

Thanks.

[Alex8211101]

Finally,I found where the problem happened,I used the "cropped‘ labels during training,I retrained and got the correct results,thanks again for your answer.

[Nic-Ma]

Cool, glad to see your progress.