3D U-Net pytorch

Hi, I’m fairly new to this field but I had some experiences with deep learning so I thought I’m going to give this a try.
When I see the data, my first thought was using 3D U-Net since the data we processing is in 3D matrix representation. My colab notebook is here. Using 3D U-Net without any significant modification and no pre-processing, I achieved 0.3 F1 score and 0.539 Accuracy.

My pipeline is:

  • Zero padding the test set into divide-able cube
  • Divide the data into cubes
  • Train with 3D U-Net
  • Validate each test cubes then stack them together into final prediction cube

Some thought and future improvement

  • No data augmentation was used
  • Pre processing method such as filtering, normalizing or thresholding was not used
  • The training was quite time-consuming at approx 30min/epoch, Im running this on colab pro

I’ll try to improve this and I would be very happy to receive any suggestion, thank you.