How to improve dice score with Lits

mist-medical / MIST

MIST: A simple, scalable, and end-to-end framework for 3D medical imaging segmentation.

Apache License 2.0

28 stars 11 forks source link

How to improve dice score with Lits #5

Closed rachelglenn closed 8 months ago

rachelglenn commented 1 year ago

I ran the tensorflow code with the Lits data with the flags

nnunet --pocket --xla -amp --patch-size 128 128 128 epochs 125

and I got the following results

Mean, 0.8993201019273616 Std, 0.1050104076908201 25th Percentile, 0.8935995292553426 Median, 0.9303144316101752 75th Percentile, 0.9463229317287942

How can I improve the mean dice score?

Thanks,

Rachel

aecelaya commented 1 year ago

I'd recommend more epochs and also try postprocessing the predictions by taking the largest component. I'll try these and a couple of other things and get back to you. Are you running a five fold cross validation?