Problem description:

• Cells labeled with red, green or both markers appear in 2-channel images
• Degree of bleed-through across imaging channels is different for each animal

Training dataset:

• Approximate cell centers are marked in each channel
• Size of a single channel 2048 x 2048 image is 32.0 Mb
• (both channels for images + labels) for a 20 images = ~1.25 Gb
• Load it all in cpu memory for patch generation while training

.
├── pwgt                          # create pixel-wise ground truth from manual cell center annotations
│   ├── script_00_optim_centers.m # remove duplicate annotations and optimize cell center position
│   ├── script_01_FM.m            # generate .mat files with output from fast marching based segmentation algorithm
│   ├── script_02_lbl_from_FM.m   # use fast marching output to create channel-wise, pixel-wise label maps  
│   ├── FM.m                      # fast marching core - output is a .mat file per image tile
│   ├── lbl_from_FM.m             # create pixelwise labels from S, T, D maps generated by fast marching
│   ├── optim_centers.m           # optimization routine to re-position manual annotations
│   ├── rem_duplicates.m          # remove redundant annotations for the same cell
│   ├── simple3d.m                # evaluates if pixel is simple or not (digital topology)
│   └── tifvol2mat.m
├── models.py
├── utils
│    ├── analysis.py
│    ├── data.py
│    ├── transforms.py
└── scripts 
    ├── pred_Ai224_RG_unet_zarr.py
    └── train_Ai224_RG_unet.py

Image format processing

utils.py contains helper functions for following format conversions:

• .jp2 (uint16) --> np.array (uint8)
• .jp2 (uint16) --> .zarr (uint16)

Notes:

• Use pip install -e . to enable importing of functions as from cell_count.utils.data import function as f
• trained models are here: dropbox link.
• commit 3e952: Input is scaled, equal weights used for cross-entropy loss. 60,000 epoch experiment
• Pixel-wise ground truth codes adapted from (topo-preserve-fastmarching repository)[https://github.com/rhngla/topo-preserve-fastmarching]
• commit 0172c: model was trained on control_retraining dataset (in addition to the initial dataset)
• trained models for this are in: /control_retraining_combined. evaluate 28960_ckpt.pt and 45000_ckpt.pt

Prediction of cell centers on full brain images

• Each slice takes ~500 s just for I/O from .zarr format.
• ~150 slices per brain will require ~1250 min for I/O
• Passing through GPU + morphological ops to get co-ordinates is negligible in comparison

References

• Commonly used loss functions for segmentations