weihuang527/pytorch_mala

pytorch_mala

Introduction:
This code referenced from funkey/mala. Paper is Large Scale Image Segmentation with Structured Loss Based Deep Learning for Connectome Reconstruction

0. Environment
   docker image: renwu527/auto-emseg:v4.1

1. Preprocessing
   Convert segmentation label to boundary
   See the cremia.py from line 106 to line 109
   GrowBoundary is to make the boundary of the label thicker
   seg_to_affgraph is to convert label to boundary
   For the different training data, we must change the name of raw image and labels in the line 86 and 102 (in the cremia.py). For different data formats (png, 3d tiff, hdf and so on), we need to change the way of reading slightly.

2. Training

   python main.py -dp='../data'

   -dp is the path of the training data, including raw image and labels.

3. Inference

   python inference_mask_16bit_para_v2.py -in='../data/***.tif' -out='../affs' -ite=100000

   -in is the path of the data need to be predicted.
   -out is the outpath of predicted affinity and mask.
   -ite is the iteration times of trained model from the step 2.

4. Over-segmentation and agglomeration

   python2 aff2id_mask_16bit_para.py -in='../affs' -out='../inference' -nu=500 -t=0.5

   -nu is the depth of inference data (the z dimension).
   -t is the threshold used to agglomerate.

5. Evaluate

   python evaluate.py -seg='../inference/segmentation.hdf' -gt='../data/gt.hdf'

   -seg is the output of the step 4.
   -gt is the groundtruth of inference data.