Assignment of Labels by 3D CRF applied on Deep Medic Output

[stalhabukhari]

Hi! I applied your 3D CRF on an output computed via Deep Medic, and used a simple Dice Overlap to compute the score.

As shown above, the dice score of Label 1 of CRF-ed output (0.006999..) is very low and hence boggling.

One question that pops into mind is, how is the CRF deciding on the output labels, given we only provide it probability maps? Parameters of my configuration file are as below:

` -numberOfModalitiesAndFiles 4 /home/talha/Desktop/CRF + Dice Score/dataset/HGG/Brats18_2013_5_1/Brats18_2013_5_1_t1_normalized.nii.gz /home/talha/Desktop/CRF + Dice Score/dataset/HGG/Brats18_2013_5_1/Brats18_2013_5_1_t1ce_normalized.nii.gz /home/talha/Desktop/CRF + Dice Score/dataset/HGG/Brats18_2013_5_1/Brats18_2013_5_1_t2_normalized.nii.gz /home/talha/Desktop/CRF + Dice Score/dataset/HGG/Brats18_2013_5_1/Brats18_2013_5_1_flair_normalized.nii.gz

-numberOfForegroundClassesAndProbMapFiles 4 /home/talha/Desktop/CRF + Dice Score/prob_maps/Brats18_2013_5_1_ProbMapClass1.nii.gz /home/talha/Desktop/CRF + Dice Score/prob_maps/Brats18_2013_5_1_ProbMapClass2.nii.gz /home/talha/Desktop/CRF + Dice Score/prob_maps/Brats18_2013_5_1_ProbMapClass3.nii.gz /home/talha/Desktop/CRF + Dice Score/prob_maps/Brats18_2013_5_1_ProbMapClass4.nii.gz

-imageDimensions 3.0 240.0 240.0 155.0

-minMaxIntensities -3.5100016593933105 9.68246078491211

-outputFolder /home/talha/Desktop/CRF + Dice Score/crf_out/HGG/Brats18_2013_5_1

-prefixForOutputSegmentationMap Brats18_2013_5_1_prediction

-prefixForOutputProbabilityMaps Brats18_2013_5_1map

-pRCZandW 3.0 3.0 3.0 3.0

-bRCZandW 17.0 12.0 10.0 5.0

-bMods 3.5 3.5 3.5 3.5

-numberOfIterations 5`

Additionally, I would like to inquire about the specific Merging technique utilized to merge the multi-class output of DeepMedic, to be used with CRF, as mentioned in the paper:

For the multi-class problem it is challenging to find a global set of parameters for the CRF which can consistently improve the segmentation of all classes. So instead we merge the four predicted probability maps into a single “whole tumour”map for CRF post-processing.

[chengjianhong]

hi @stalhabukhari , do you success to use it ?

[stalhabukhari]

Hi @chengjianhong, I am really sorry for such a late reply, don't know how I missed this. Yes, it works like a charm. I just compiled it again, don't know why it wasn't working previously. I should close this issue as well.