Usiigaci

We have created Usiigaci based on the Mask R-CNN repository from Matterport for label-free instance-aware segmentation of cells under phase contrast microscopy.

The results were astonishing. 

We further developed tracking module and data processing module for Usiigaci.

Now Usiigaci is an all-in-one solution for segmentation, tracking, and analysis for cellular migration with accurate resolution to cellular outline. 
Moreover, with the GUI with the tracking module, users are allowed to review the segmentation and tracking results and revoke bad tracks from segmentation or tracking or biologically invalid cell tracks. This opportunity of manual intervention was not equipped in many software but we believe a manual intervention for data verification is essential to ensure the validity of single cell tracking data.

The paper is submitted to SoftwareX. (20190225 Accepted)
The sourcecode of Usiigaci is released with MIT license and you can find it in our github repo.



https://lilianweng.github.io/lil-log/2017/12/31/object-recognition-for-dummies-part-3.html


regional convolutional neural network is another great improvement for image segmentation and winner of Marr prize at ICCV 2017 , but Kaiming He et al at Facebook A Research.

By instance segmentation, individual target ROI can be recognized from the feature map through regional proposal network and roi aligned to get accurate mask. 
this way, even rois with the same class are together as in the case for semantic segmentation, in whichi no separation of rois can be made, in instance aware segmentation of mask-rcnn, individual instance can be segmented. 
As in cell microscopy's case, individual cells may touch each other, and separation is crucial. if the masks are touched together and cannot be separated, then there is no way to analyze it.

In cell microscopy, mask-rcnn architecture is especially useful.



Environment description:
python 3.6
TensorFlow 1.9 (1.7 on linux)
Keras 
Geforce GTX 1070 mobile driver verson: with CUDA 9.0 CuDNN 
total memory 7.92GB free memory 7.65GB
1920 cuda cores 1.645GHz
prediction 1 model 1.87s/it (per image)






if using Geforce GTX1080Ti Desktop CUDA9.0
3584 cuda cores 1.582GHz
total memory  11GB


Tips on running on windows 

set "KERAS_BACKEND=tensorflow"

install gpu enabled tensorflow on ananconda 
conda install -c anancoda tensorflow-gpu

check on tensorflow gpu availability
from tensorflow.python.client import device_lib
print(device_lib.list_local_devices())