Wolfram Cloud Document

A few years ago I had shared a paclet for making UNET convolutional neural networks to this community. Although, I stopped developing the stand alone paclet since I incorporated this in my main paclet QMRITools (QMRITools.com) i did not stopped developing the UNet functionality. It is sill being used to improve automated segmentation of leg muscles which is done by the function SegementData which follows the above flowchart

In mathematica it looks like this

Over the years I have experimented with many different UNET architectures which can all be generated with the function MakeUnet which has changed a lot since the original UNET paclet. Below is a demonstration of its current capabilities in the latest version of QMRITools.

<<QMRITools`netIm=Information[#,"SummaryGraphic"]&;makeNetIm=With[{im=Information[#,"SummaryGraphic"]},Show[im,AspectRatio->0.5,ImageSize->Max[AbsoluteOptions[im,ImageSize][[1,2]]]]]&;netSumIm=Row[{NetSummary[#],netIm[#]}]&;

MakeUnet Function

In[]:=

Echo[Grid@Transpose@{Options@MakeUnet},"MakeUnet options"];Echo[Show[MakeFunctionGraph[MakeUnet,LabelPlacement->Automatic,AllowSelfDependencies->True],ImageSize->600],"Function graph"];Echo[Row[{Darker@Green,"Public function ",Blue,"Private function"}],"Legend"];

MakeUnet options

NetworkArchitectureUNet

BlockTypeResNet

ActivationTypeGELU

RescaleMethodConv

NetworkDepth5

DownsampleScheduleAutomatic

SettingScheduleAutomatic

FeatureSchedule32

DropoutRate0.2

MonitorCalcFalse

Function graph

Legend

Public function

Private function

The internal Make node function

MakeNode with options

In[]:=

scale={2,2};(*{inputscaling,outputscaling}*)skipConn={1,True};(*{numberofinputskipconnetions,outputskipconnnectionyes/no}*)conf={"Conv",{32,16},{"GELU",3}};(*nodeconfiguration*)

In[]:=