Applying metrics

[Meddebma]

Hi,

I was evaluating the unet for spleen segmentation with different metrics, but I got en Error trying to compute the ROC:

here is the code:
with torch.no_grad():
metric_sum = 0.0
metric_count = 0
for val_data in val_loader:
# define sliding window size and batch size for windows inference
val_images, val_labels = val_data["image"].to(device), val_data["label"].to(device)
roi_size = (96, 96, 96)
sw_batch_size = 4
val_outputs = sliding_window_inference(val_data["image"].to(device), roi_size, sw_batch_size, model)
val_outputs = post_trans(val_outputs)
value = compute_meandice(y_pred=val_outputs, y=val_labels, include_background=False)
roc = compute_roc_auc(y_pred=val_outputs, y=val_labels)
conf = get_confusion_matrix(y_pred=val_outputs, y=val_labels, include_background=False)
hausdorff = compute_hausdorff_distance(y_pred=val_outputs, y=val_labels, include_background=False)
metric_count += len(value)
metric_sum += value.item() * len(value)
#print(f"val_loss:", loss)
print(f"val_dice:", value)
print(f"Confusion Matrix:", conf)
print(f"Hausdorff:",hausdorff)
print(f"roc_auc:", roc)

and here is the Error:

ValueError Traceback (most recent call last)

in ()
30 y=val_labels,
31 include_background=False)
---> 32 roc = compute_roc_auc(y_pred=val_outputs, y=val_labels)
33 conf = get_confusion_matrix(y_pred=val_outputs, y=val_labels, include_background=False)
34 hausdorff = compute_hausdorff_distance(y_pred=val_outputs, y=val_labels, include_background=False)

/usr/local/lib/python3.7/dist-packages/monai/metrics/rocauc.py in compute_roc_auc(y_pred, y, to_onehot_y, softmax, other_act, average)
100 y_ndim = y.ndimension()
101 if y_pred_ndim not in (1, 2):
--> 102 raise ValueError("Predictions should be of shape (batch_size, n_classes) or (batch_size, ).")
103 if y_ndim not in (1, 2):
104 raise ValueError("Targets should be of shape (batch_size, n_classes) or (batch_size, ).")

ValueError: Predictions should be of shape (batch_size, n_classes) or (batch_size, ).

How should I fix this?

Best, Aymen

[rijobro]

As the error indicates, the y_pred of compute_roc_auc is expected to be a 1- or 2D array with the predicted classes. In your case, however, you are passing your segmented image.

Some metrics make sense for classification but don't make sense for segmentation. ROC/AUC and confusion matrices are examples of this (e.g., see this blog post).

[Meddebma]

Thanks for the reading, this makes totally sense!