More details of the baseline

Hi,

I think the provided baseline model is actually quite good, and I'm curious about how you trained it. Is it possible that you can give more training details of the baseline model? Thank you.

Hi,

We first have oversampled the positive cases to reach the number of negative cases (so that they both are represented equally during training). Then we have trained the model for 20 epochs, and chose the model with the best AUC score on the validation set. Our validation set was randomly selected from training set and contains around ~200 images.

Let us know if you would like to know anything more about it,

Best, Ecem

Hi Ecem,

Thank you for the reply. Is the learning strategy the same as the baseline code (batch size=2, SGD, init lr=0.005, decay every 3 epochs)? And any other modifications such as augmentation, anchors, training image size? Thank you.

Best

By the way, the AUC score on validation set seems to be quite good (over 0.94) in my case. I wonder if it is the same to you? Thank you.

Hi,

The training strategy is the same as used in the baseline code. We did not apply any other changes.

Indeed, AUC score in the validation set also goes up to 0.95 for the baseline model, but the secret test data contains images that are a bit more difficult as they were annotated based on the CT scans. So, the baseline method achieved a lower performance there.

Best, Ecem