How to calculate 3D data Dice score via 2D network inference?

[Mentholatum]

I am having trouble when using monai.metrics.DiceMetric. I want to calculate 3d data dice score via 2d network prediction. Here is my code:

    dice_acc = monai.metrics.DiceMetric(include_background=True,
                          reduction=MetricReduction.MEAN,
                          get_not_nans=True)
    post_label = monai.transforms.AsDiscrete(to_onehot=args.classes)
    post_pred = monai.transforms.AsDiscrete(argmax=True,
                           to_onehot=args.classes)
    with torch.no_grad():
        for idx, input_tuple in enumerate(val_generator):
            data, target = prepare_input(input_tuple=input_tuple, args=args)
            # data and target shape : [1, 1, 32, 256, 256]
            prediction = torch.zeros_like(target)
            for ind in range(data.shape[2]): # prediction by slice
                slice = data[:, :, ind, :, :]
                gt_slice = target[:, :, ind, :, :]
                model.eval()
                with torch.no_grad():
                    outputs = model(slice)
                    # **confused about this part**
                    val_labels_list = decollate_batch(gt_slice)
                    val_labels_convert = [post_label(val_label_tensor) for val_label_tensor in val_labels_list]
                    val_outputs_list = decollate_batch(outputs)
                    val_output_convert = [post_pred(val_pred_tensor) for val_pred_tensor in val_outputs_list]
                    # calculate slice result
                    # Dice
                    dice_acc(y_pred=val_output_convert, y=val_labels_convert)
                    out = torch.argmax(torch.softmax(outputs, dim=1), dim=1)
                    pred = out
                    prediction[0, 0, ind] = pred

            val_labels_list = decollate_batch(target)
            val_labels_convert = [post_label(val_label_tensor) for val_label_tensor in val_labels_list]
            val_outputs_list = decollate_batch(prediction)
            val_output_convert = [post_pred(val_pred_tensor) for val_pred_tensor in val_outputs_list]

            # Dice
            dice_acc(y_pred=val_output_convert, y=val_labels_convert)
            acc1 = dice_acc.aggregate()
            avg_acc1 = acc1[0].detach().cpu().numpy()
            dice += avg_acc1
        print('DSC:', format(float(dice) / len(val_generator), '.4f'))

when I use dice_acc() to get slice score, the final volume data dice seems like the slice dice score average. But when I delete the dice_acc() function during slice inference process, the final result always seems to be close to 0.5, and the results do not change with other models. How should I use this function correctly? Thank you very much for answering my question!

[Mentholatum]

I think the problem is my post_label and post_pred function. I use argmax twice so that I alway get the fixed wrong anwser.