improve model performance

[bodokaiser]

One

Architecture

L1 Loss L2 Loss

Plots

Testing

Training

Two

Architecture

L1 Loss L2 Loss

Plots

Testing

Training

U-Net

Architecture

L1 Loss L2 Loss

Plots

Testing

Training

[bodokaiser]

U-Net (lr=1e-4)

L1 Loss L2 Loss

Plots

Testing

Training

[bodokaiser]

@albarqouni I started run with lr=1e-5. It should be finished tomorrow. If you want to check image progress over epochs you can visit http://52.15.106.125:3000/ until tomorrow.

[bodokaiser]

U-Net (lr=1e-5)

L1 Loss L2 Loss

Plots

Testing

Training

[albarqouni]

What kind of pre-processing you have? Normalization ... etc? Do you have any batch-normalization? Have you checked the gradient histogram?

[bodokaiser]

I apply these transforms:

1. input images (MR)

   • clipping ([–32 768, 32 767] -> [-32 000, 0] -> [0, 32 000])
   • histnorm (nbins=256)

2. target images (US)

   • histnorm (nbins=256)

I use skimage.exposure.equalize_hist for histogram normalization.

[bodokaiser]

No, that u-net model has no BatchNorm layers (it's as close as possible to the original one).

What do you mean with gradient histogram?

[albarqouni]

I wouldn't do any histogram equalization at the moment. You might need to have a Batch-Normalization layer before the first convolutional layer. Hopefully, that helps!

[bodokaiser]

What about having BatchNorm after every Conv layer?

[albarqouni]

That's possible. However, we don't wanna change the architecture at the moment. The first layer acts as a pre-processing step.

[bodokaiser]

U-Net

• lr=1e-5
• no histogram normalization
• batch normalization before first convolution

L1 Loss L2 Loss

Plots

Testing

Training