Compute SurfaceDice

[Yangtting]

Hello, I'm using SurfaceDiceMetric in validation stage in my auto-segmentation project. However, I can't understand the 'class_thresholds', thus I don't know how to use it in my code. Here is my code. Could you please help me?

post_transforms = Compose(
    [
        Invertd(
            keys="pred",
            transform=val_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,
            device="cpu",
        ),
        AsDiscreted(keys="pred", argmax=True, to_onehot=2),
        AsDiscreted(keys="label", to_onehot=2),
        # SaveImaged(keys="pred", meta_keys="pred_meta_dict", output_dir="./out", output_postfix="seg", resample=False),
    ]
)

surfacedice_metric=SurfaceDiceMetric(class_thresholds=[0.9],include_background=False, reduction="mean")

model.load_state_dict(torch.load(os.path.join(root_dir, "best_metric_model.pth")))
model.eval()
i=1
with torch.no_grad():
    for val_data in val_loader:
        val_inputs = val_data["image"].to(device)
        roi_size = (160, 160, 160)
        sw_batch_size = 4
        val_data["pred"] = sliding_window_inference(val_inputs, roi_size, sw_batch_size, model, overlap=0.8)
        val_data = [post_transforms(i) for i in decollate_batch(val_data)]
        val_outputs, val_labels = from_engine(["pred", "label"])(val_data)
        # compute metric for current iteration
        dice_metric(y_pred=val_outputs, y=val_labels)
        surfacedice_metric(y_pred=val_outputs, y=val_labels)
    # aggregate the final mean dice result
    dice_metric_org = dice_metric.aggregate().item()
    surfacedice_metric_org=surfacedice_metric.aggregate().item()
    # reset the status for next validation round
    dice_metric.reset()
    surfacedice_metric.reset()

print("Metric on original image spacing: ", dice_metric_org)
print("Metric on original image spacing: ", surfacedice_metric_org)

[KumoLiu]

Hi @Yangtting, SurfaceDiceMetric determines which fraction of a segmentation boundary is correctly predicted. A boundary element is considered correctly predicted if the closest distance to the reference boundary is smaller than or equal to the specified threshold (class_thresholds), which is related to the (clinically) acceptable amount of deviation in pixels.
It should be a list of class-specific thresholds in pixels.
https://github.com/Project-MONAI/MONAI/blob/8e134b8cb92e3c624b23d4d10c5d4596bb5b9d9b/monai/metrics/surface_dice.py#L44