Datastore Mask R-CNN for kidney segmentation

I have

• several CT images 512x512 with normalized values between 0 and 1. Basically 2dimensional black white images
• Kidney labels for each images with 0 or 1
• for each image 2 boxes with x,y,height, width (for each kidney of the human)
• and the labels I could generate (always 'kidney' I guess?)

How do I generate a datastore to train the network, as described here under "Create Datastore":

https://de.mathworks.com/help/vision/ug/example-InstanceSegmentationUsingMaskRCNNDeepLearningExample.html#mw_rtc_PerformInstanceSegmentationUsingMaskRCNNExample_M_B1AB0600

They convert it somehow from some MAT files with some custom function, hence I do not know how to assemble my own training data into a data store

Answer 1

[Prince Kumar]

Hi,

Please read the "Create Datastore that Reads Data" section from the following documentation.

• https://www.mathworks.com/help/vision/ug/getting-started-with-mask-r-cnn-for-instance-segmentation.html

You can also read Datastores for deep learning for more understanding.

Hope this helps!

[Odo Luo]

I have trouble understanding the needed sizes. I managed to put up a datastore but got following error:

Error using trainMaskRCNN
[net,info] = trainMaskRCNN(ds,net,options,FreezeSubNetwork="backbone");
↑
Invalid argument at position 1. Expected input number 1, trainingData, to be of size 1x4, but it is of size 97x4.

Error in maskRcnnTest (line 25)
[net,info] = trainMaskRCNN(ds,net,options,FreezeSubNetwork="backbone");

As 1 CT Image consists of several images/slides and I used 1 CT image with 97 slieds, I have a 97x4 ds.

Nevertheless trainMaskRCNN expects 1x4 ? How is a network supposed to train on 1 image alone ?

Part of my ds preview:

ans =

97×4 cell array

{512×512 double}    {0×0 double    }    {0×0 double}    {512×512×2 logical}
{512×512 double}    {0×0 double    }    {0×0 double}    {512×512×2 logical}
{512×512 double}    {0×0 double    }    {0×0 double}    {512×512×2 logical}
{512×512 double}    {0×0 double    }    {0×0 double}    {512×512×2 logical}
{512×512 double}    {0×0 double    }    {0×0 double}    {512×512×2 logical}
{512×512 double}    {[52 226 25 25]}    {'kidney'  }    {512×512×2 logical}
{512×512 double}    {[47 212 38 35]}    {'kidney'  }    {512×512×2 logical}
{512×512 double}    {2×4 double    }    {2×6 char  }    {512×512×2 logical}
{512×512 double}    {2×4 double    }    {2×6 char  }    {512×512×2 logical}
{512×512 double}    {2×4 double    }    {2×6 char  }    {512×512×2 logical}

[Prince Kumar]

I think you are not preparing the data properly.

Try to prepare the boxLabelDatastore properly. You can check the examples here : link

Then try to create combine the image datastore and box label datastore.

[Odo Luo]

Thank you for your help!

I have a more concrete question regarding the image datastore and made a seperate question!

But you put me on the right track, lots of thanks!