Purpose of the two final convolutional layers

[tsuijenk]

Good day,

Thank you for your amazing work! I have been wondering why there are two final convolutional layers, in contrast to UNet's traditional one?

1. Does it outperform UNet's one final convolutional layer?
2. I have also been wondering about the specs of these two layers. Does it go like this? Please correct me if I misinterpreted it!

• 1st final convolutional layer: 48 features -> an output head channel (through 3x3 convolutional layer): Conv -> BN -> ReLU
• 2nd final convolutional layer: an output head channel -> an output head channel (through 1x1 convolutional layer): Conv -> Sigmoid

Thank you.

[Haydnspass]

Hey @tsuijenk
adding a few numbers of final layers (called "heads") is a pretty standard procedure if you have multiple (and different) outputs.

[ASpeiser]

To expand a bit: the standard U-nets used for segmentation have just one binary output, while we have 10 output channels that also perform regression, therefore it makes sense to add some parameters at that stage.

Generally speaking though, changing details of the architecture (e.g. the exact number of layers, feature channels, nonlinearity, etc.) will not have a large impact on the performance of DECODE. The most important training hyperparameters are the learning rate and batch size. And the most important thing overall is to set up the simulation parameters correctly.

Oh, and we don't use any Batch normalization in our training. This has adverse effects, because absolute intensities are critical to optimally analyze this kind of data.