Microscopy segmentation exps

[NIRVANALAN]

• Arch: UNet
• Training data: 512*512 patches of 14 slides(resized to 574 for UNet input)
• lr=0.01, momentum=0.99,epoch=50,batch=4
• loss: Pixel CE

All data patches data split:

Trainset: 1.8k Testset: 400

Remove pure BG images(without organelles)

Train: 1444 test: 362

results on several dataset split:

settings	best ACC	best loss
1000 unfiltered	93.157772	0.292162
800 unfiltered
1000 unfiltered, lr=0.02	93.151519	0.300175

[NIRVANALAN]

Considering the BG pixels (pixels that are not labeled) dominates the images, category CE loss may lead to overfitting(all pixels are predicted as BG pixels, and the ACC is still high), this time I tried with BCE+DICE Loss and trained models without BG pixels. The output of different class lies in different channel, and softmax as final layer activation function.

Also, 3 metrics are used to evaluate the model.

1. Pixel Accuracy
2. Multi Class Pixel Accuracy
3. Multi Class IOU