Detect and track cells in bioimage timeseries.
Produce results in a variety of formats including ISBI CTC compatible.
This project is a consolidation of devseg_2 which reduces all code necessary for data generation, detection training, detection prediction, segmentation, tracking and evaluation to 1k lines of python.
-------------------------------------------------------------------------------
Language files blank comment code
-------------------------------------------------------------------------------
Python 6 324 315 1025
Markdown 1 11 0 75
Bourne Shell 3 11 39 32
Text 1 2 0 15
-------------------------------------------------------------------------------
SUM: 11 348 354 1147
-------------------------------------------------------------------------------
First, make sure you're python environment is set up: pip install -r requirements.txt.
Then download some training datasets from isbi: sh download-isbi-data.sh.
Run with python cpnet.py Fluo-C2DL-Huh7 (or any ISBI dataset name) assuming a directory structure like:
Fluo-C2DL-Huh7/
├── 01
│ ├── t000.tif
│ ├── ...
├── 01_GT
│ ├── SEG
│ │ ├── man_seg000.tif
│ │ ├── ...
│ └── TRA
│ ├── man_track.txt
│ ├── man_track000.tif
│ ├── ...
├── 02
│ ├── t000.tif
│ ├── ...
└── 02_GT
├── SEG
│ ├── man_seg000.tif
│ ├── ...
└── TRA
├── man_track.txt
├── man_track000.tif
├── ...
which produces...
cpnet-out/Fluo-C2DL-Huh7/
├── data
│ ├── dataset.pkl ## patches with `raw`, `target` and `mask`.
│ └── png
│ ├── t000-d0000.png ## time:000 - patch:0000
│ └── ...
├── predict
│ ├── t0000-latest.png ## time:0000 - model:latest
│ └── ...
├── track
│ ├── l000.png ## time:000 colored marker for each object
│ └── ...
└── train
├── glance_output_train
│ ├── a000_000.png ## time:000 patch:000 (in training set)
│ └── ...
├── glance_output_vali
│ ├── a000_000.png ## time:000 patch:000 (in validation set)
│ └── ...
├── history.pkl ## loss and validation scores over time (epochs)
├── labels.pkl ## train|vali label for each patch in `dataset.pkl`
└── m
├── best_weights_f1.pt ## best model weights according to f1 metric
├── best_weights_height.pt
├── best_weights_latest.pt
└── best_weights_loss.pt