How to calculate composite dice for multi-level segmentation instead of average dice

[Alex8211101]

When calculating the dice of organs of different sizes, I found that monai is to obtain the meandice of different organs, but the larger organs often occupy a larger pixel value. The calculation of meandice will affect the comprehensive value due to the lower dice of small organs. Is there any such method? Thank you in advance!

`dice_metric = DiceMetric(include_background=True, reduction="mean")
dice_metric_batch = DiceMetric(include_background=True, _reduction="mean_batch")

These are my output dices.
Epoch: 99, mean dice: 0.834375,femur dice: 0.949789,femoral cartilage dice: 0.736557,tibia dice:0.943555,tibia cartlage dice:0.707600

[Nic-Ma]

Hi @Alex8211101 ,

Thanks for your interest here.
If I understand your question correctly, I think you can use reduction="mean_channel" to get mean dice of every image, then compute the weighted average by image sizes.
What do you think?

Thanks.

[Nic-Ma]

Hi @Alex8211101 ,

I checked your program, seems you missed Activationsd(keys="pred", softmax=True) in your post transforms of evaluation?
Could you please help add it and try again?

Thanks.

[Alex8211101]

Thanks for the reminder,I made the modification as below.
post_transforms = Compose([ EnsureTyped(keys="pred"), Invertd( keys="pred", 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, ), Activationsd(keys='pred',softmax=True), AsDiscreted(keys="pred", argmax=True, to_onehot=2), # AsDiscreted(keys="label", to_onehot=2), ])
But unfortunately I only got the result od dice=0.14.That's weird.I don't know if it's because of my wrong setting during traing that I get the segmentation results that looks good.Thank you very much for your help!
` dice_metric = DiceMetric(include_background=False, reduction='mean',)

post_trans=Compose([EnsureType(),Activations(sigmoid=True),AsDiscrete(argmax=True,to_onehot=2)])
post_label = Compose([EnsureType(), AsDiscrete(to_onehot=2)])

loss_function = monai.losses.DiceCELoss(include_background=True, to_onehot_y=True, softmax=True)
torch.backends.cudnn.benchmark = opt.benchmark

val_interval = 1
best_metric = -1
best_metric_epoch = -1
epoch_loss_values = []
metric_values = []
save_name_txt = os.path.join(opt.output_dir, "log_dir", 'Res_log_femur.txt')
model_file = open(save_name_txt, "w")
total_start=time.time()

writer = SummaryWriter()
for epoch in range(opt.epochs):
    epoch_start = time.time()
    print("-" * 10)
    print(f"epoch {epoch + 1}/{opt.epochs}")
    net.train()
    epoch_loss = 0
    step = 0
    for batch_data in train_loader:
        step += 1
        inputs, labels = batch_data["image"].cuda(), batch_data["label"].cuda()
        optim.zero_grad()
        outputs = net(inputs)
        # outputs=inferer(inputs.cuda(),net)
        loss = loss_function(outputs, labels)
        loss.backward()
        optim.step()
        epoch_loss += loss.item()
        epoch_len = len(check_train) // train_loader.batch_size
        print(f"{step}/{epoch_len}, train_loss: {loss.item():.4f}")
        writer.add_scalar("train_loss", loss.item(), epoch_len * epoch + step)
    epoch_loss /= step
    epoch_loss_values.append(epoch_loss)
    print(f"epoch {epoch + 1} average loss: {epoch_loss:.4f}")

    train_result = 'Epoch: %d, Train Loss: %f' % (epoch, epoch_loss)

    if (epoch + 1) % val_interval == 0:
        net.eval()
        with torch.no_grad():
            for val_data in val_loader:
                val_inputs, val_labels = (
                    val_data["image"].cuda(),
                    val_data["label"].cuda(),
                )
                roi_size = (256, 256, 48)
                sw_batch_size = 4
                val_outputs = sliding_window_inference(
                    val_inputs, roi_size, sw_batch_size, net)
                val_outputs = [post_trans(i) for i in decollate_batch(val_outputs)]
                val_labels = [post_label(i) for i in decollate_batch(val_labels)]
                # compute metric for current iteration
                dice_metric(y_pred=val_outputs, y=val_labels)


                # aggregate the final mean dice result
            metric = dice_metric.aggregate().item()
            # reset the status for next validation round
            dice_metric.reset()
            # dice_metric_batch.reset()
            torch.save(net.state_dict(), os.path.join(
                opt.root_dir, 'resUNet3d_femur',
                "{}_Dice:{:.4f}.pth".format(epoch, metric)))
            metric_values.append(metric)
            if metric > best_metric:
                best_metric = metric
                best_metric_epoch = epoch + 1
                torch.save(net.state_dict(), os.path.join(
                    'log/resUNet3d_femur', "best_metric_model.pth"))
                print("saved new best metric model")
            print(
                f"current epoch: {epoch + 1} current mean dice: {metric:.4f}" 
                f"\nbest mean dice: {best_metric:.4f} "
                f"at epoch: {best_metric_epoch}"
            )
            valid_result = 'Epoch: %d, mean dice: %4f' % (
                epoch, metric)
    model_file.write(valid_result + '\n')
    model_file.write(train_result + '\n')
    print(f"time consuming of epoch{epoch + 1} is: {(time.time() - epoch_start):.4f} s")
total_time = time.time() - total_start
model_file.close()`

[Nic-Ma]

Please put the Activationsd before Invertd, just like your inference program.

Thanks.

[Alex8211101]

Thanks again for your reply, the monai team has been so helpful to me. I can already export the results of the segmentation, but there is one strange place. I input the best metric model for training the knee tibia, but the output turned out to be the segmentation result of the femur, and the accuracy of the segmentation is relatively high. I can confirm the data is accurate through the two-dimensional result of check best model.

[Nic-Ma]

I didn't get your problem "I input the best metric model for training the knee tibia, but the output turned out to be the segmentation result of the femur".
Maybe there is some confusion in your jupyter notebook? Could you please help totally restart your jupyter notebook and try again?

Thanks.